diff options
Diffstat (limited to 'megapixels/commands/processor')
18 files changed, 2434 insertions, 0 deletions
diff --git a/megapixels/commands/processor/_old_files_to_face_rois.py b/megapixels/commands/processor/_old_files_to_face_rois.py new file mode 100644 index 00000000..d92cbd74 --- /dev/null +++ b/megapixels/commands/processor/_old_files_to_face_rois.py @@ -0,0 +1,168 @@ +""" +Crop images to prepare for training +""" + +import click +# from PIL import Image, ImageOps, ImageFilter, ImageDraw + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +color_filters = {'color': 1, 'gray': 2, 'all': 3} + +@click.command() +@click.option('-i', '--input', 'opt_fp_files', required=True, + help='Input file meta CSV') +@click.option('-o', '--output', 'opt_fp_out', required=True, + help='Output CSV') +@click.option('-e', '--ext', 'opt_ext', + default='jpg', type=click.Choice(['jpg', 'png']), + help='File glob ext') +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('-t', '--detector-type', 'opt_detector_type', + type=cfg.FaceDetectNetVar, + default=click_utils.get_default(types.FaceDetectNet.DLIB_CNN), + help=click_utils.show_help(types.FaceDetectNet)) +@click.option('-g', '--gpu', 'opt_gpu', default=0, + help='GPU index') +@click.option('--conf', 'opt_conf_thresh', default=0.85, type=click.FloatRange(0,1), + help='Confidence minimum threshold') +@click.option('-p', '--pyramids', 'opt_pyramids', default=0, type=click.IntRange(0,4), + help='Number pyramids to upscale for DLIB detectors') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('--display/--no-display', 'opt_display', is_flag=True, default=False, + help='Display detections to debug') +@click.option('--recursive/--no-recursive', 'opt_recursive', is_flag=True, default=False, + help='Use glob recursion (slower)') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('--color', 'opt_color_filter', + type=click.Choice(color_filters.keys()), default='color', + help='Filter to keep color or grayscale images (color = keep color') +@click.pass_context +def cli(ctx, opt_dirs_in, opt_fp_out, opt_ext, opt_size, opt_detector_type, + opt_gpu, opt_conf_thresh, opt_pyramids, opt_slice, opt_display, opt_recursive, opt_force, opt_color_filter): + """Converts frames with faces to CSV of ROIs""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import dlib # must keep a local reference for dlib + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils + from app.processors import face_detector + + # ------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + + if not opt_force and Path(opt_fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + if opt_detector_type == types.FaceDetectNet.CVDNN: + detector = face_detector.DetectorCVDNN() + elif opt_detector_type == types.FaceDetectNet.DLIB_CNN: + detector = face_detector.DetectorDLIBCNN(opt_gpu) + elif opt_detector_type == types.FaceDetectNet.DLIB_HOG: + detector = face_detector.DetectorDLIBHOG() + elif opt_detector_type == types.FaceDetectNet.MTCNN: + detector = face_detector.DetectorMTCNN() + elif opt_detector_type == types.FaceDetectNet.HAAR: + log.error('{} not yet implemented'.format(opt_detector_type.name)) + return + + + # ------------------------------------------------- + # process here + color_filter = color_filters[opt_color_filter] + + # get list of files to process + fp_ims = [] + for opt_dir_in in opt_dirs_in: + if opt_recursive: + fp_glob = join(opt_dir_in, '**/*.{}'.format(opt_ext)) + fp_ims += glob(fp_glob, recursive=True) + else: + fp_glob = join(opt_dir_in, '*.{}'.format(opt_ext)) + fp_ims += glob(fp_glob) + log.debug(fp_glob) + + + if opt_slice: + fp_ims = fp_ims[opt_slice[0]:opt_slice[1]] + log.debug('processing {:,} files'.format(len(fp_ims))) + + + data = [] + + for fp_im in tqdm(fp_ims): + im = cv.imread(fp_im) + + # filter out color or grayscale iamges + if color_filter != color_filters['all']: + try: + is_gray = im_utils.is_grayscale(im) + if is_gray and color_filter != color_filters['gray']: + log.debug('Skipping grayscale image: {}'.format(fp_im)) + continue + except Exception as e: + log.error('Could not check grayscale: {}'.format(fp_im)) + continue + + try: + bboxes = detector.detect(im, opt_size=opt_size, opt_pyramids=opt_pyramids) + except Exception as e: + log.error('could not detect: {}'.format(fp_im)) + log.error('{}'.format(e)) + fpp_im = Path(fp_im) + subdir = str(fpp_im.parent.relative_to(opt_dir_in)) + + for bbox in bboxes: + # log.debug('is square: {}'.format(bbox.w == bbox.h)) + nw,nh = int(bbox.w * im.shape[1]), int(bbox.h * im.shape[0]) + roi = { + 'fn': fpp_im.stem, + 'ext': fpp_im.suffix.replace('.',''), + 'x': bbox.x, + 'y': bbox.y, + 'w': bbox.w, + 'h': bbox.h, + 'image_height': im.shape[0], + 'image_width': im.shape[1], + 'subdir': subdir} + bbox_dim = bbox.to_dim(im.shape[:2][::-1]) # w,h + data.append(roi) + + # debug display + if opt_display and len(bboxes): + im_md = im_utils.resize(im, width=min(1200, opt_size[0])) + for bbox in bboxes: + bbox_dim = bbox.to_dim(im_md.shape[:2][::-1]) + cv.rectangle(im_md, bbox_dim.pt_tl, bbox_dim.pt_br, (0,255,0), 3) + cv.imshow('', im_md) + while True: + k = cv.waitKey(1) & 0xFF + if k == 27 or k == ord('q'): # ESC + cv.destroyAllWindows() + sys.exit() + elif k != 255: + # any key to continue + break + + # save date + file_utils.mkdirs(opt_fp_out) + df = pd.DataFrame.from_dict(data) + df.to_csv(opt_fp_out, index=False)
\ No newline at end of file diff --git a/megapixels/commands/processor/cluster.py b/megapixels/commands/processor/cluster.py new file mode 100644 index 00000000..419091a0 --- /dev/null +++ b/megapixels/commands/processor/cluster.py @@ -0,0 +1,47 @@ +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg +from app.utils.logger_utils import Logger + +@click.command() +@click.option('--data_store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.NAS), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--metadata', 'opt_metadata', required=True, + type=cfg.MetadataVar, + show_default=True, + help=click_utils.show_help(types.Metadata)) +@click.pass_context +def cli(ctx, opt_data_store, opt_dataset, opt_metadata): + """Display image info""" + + # cluster the embeddings + print("[INFO] clustering...") + clt = DBSCAN(metric="euclidean", n_jobs=args["jobs"]) + clt.fit(encodings) + + # determine the total number of unique faces found in the dataset + labelIDs = np.unique(clt.labels_) + numUniqueFaces = len(np.where(labelIDs > -1)[0]) + print("[INFO] # unique faces: {}".format(numUniqueFaces)) + # load and display image + im = cv.imread(fp_im) + cv.imshow('', im) + + while True: + k = cv.waitKey(1) & 0xFF + if k == 27 or k == ord('q'): # ESC + cv.destroyAllWindows() + sys.exit() + elif k != 255: + # any key to continue + break
\ No newline at end of file diff --git a/megapixels/commands/processor/crop.py b/megapixels/commands/processor/crop.py new file mode 100644 index 00000000..778be0c4 --- /dev/null +++ b/megapixels/commands/processor/crop.py @@ -0,0 +1,104 @@ +""" +Crop images to prepare for training +""" + +import click +from PIL import Image, ImageOps, ImageFilter, ImageDraw + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +@click.command() +@click.option('-i', '--input', 'opt_dir_in', required=True, + help='Input directory') +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory') +@click.option('-e', '--ext', 'opt_ext', + default='jpg', type=click.Choice(['jpg', 'png']), + help='File glob ext') +@click.option('--size', 'opt_size', + type=(int, int), default=(256, 256), + help='Output image size') +@click.option('-t', '--crop-type', 'opt_crop_type', + default='center', type=click.Choice(['center', 'mirror', 'face', 'person', 'none']), + help='Force fit image center location') +@click.pass_context +def cli(ctx, opt_dir_in, opt_dir_out, opt_ext, opt_size, opt_crop_type): + """Crop, mirror images""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + from tqdm import tqdm + + + from app.utils import logger_utils, file_utils, im_utils + + # ------------------------------------------------- + # process here + + log = logger_utils.Logger.getLogger() + log.info('crop images') + + # get list of files to process + fp_ims = glob(join(opt_dir_in, '*.{}'.format(opt_ext))) + log.debug('files: {}'.format(len(fp_ims))) + + # ensure output dir exists + file_utils.mkdirs(opt_dir_out) + + for fp_im in tqdm(fp_ims): + im = process_crop(fp_im, opt_size, opt_crop_type) + fp_out = join(opt_dir_out, Path(fp_im).name) + im.save(fp_out) + + +def process_crop(fp_im, opt_size, crop_type): + im = Image.open(fp_im) + if crop_type == 'center': + im = crop_square_fit(im, opt_size) + elif crop_type == 'mirror': + im = mirror_crop_square(im, opt_size) + return im + +def crop_square_fit(im, size, center=(0.5, 0.5)): + return ImageOps.fit(im, size, method=Image.BICUBIC, centering=center) + +def mirror_crop_square(im, size): + # force to even dims + if im.size[0] % 2 or im.size[1] % 2: + im = ImageOps.fit(im, ((im.size[0] // 2) * 2, (im.size[1] // 2) * 2)) + + # create new square image + min_size, max_size = (min(im.size), max(im.size)) + orig_w, orig_h = im.size + margin = (max_size - min_size) // 2 + w, h = (max_size, max_size) + im_new = Image.new('RGB', (w, h), color=(0, 0, 0)) + + #crop (l, t, r, b) + if orig_w > orig_h: + # landscape, mirror expand T/B + im_top = ImageOps.mirror(im.crop((0, 0, margin, w))) + im_bot = ImageOps.mirror(im.crop((orig_h - margin, 0, orig_h, w))) + im_new.paste(im_top, (0, 0)) + im_new.paste(im, (margin, 0, orig_h + margin, w)) + im_new.paste(im_bot, (h - margin, 0)) + elif orig_h > orig_w: + # portrait, mirror expand L/R + im_left = ImageOps.mirror(im.crop((0, 0, margin, h))) + im_right = ImageOps.mirror(im.crop((orig_w - margin, 0, orig_w, h))) + im_new.paste(im_left, (0, 0)) + im_new.paste(im, (margin, 0, orig_w + margin, h)) + im_new.paste(im_right, (w - margin, 0)) + + return im_new.resize(size) + + +def center_crop_face(): + pass + +def center_crop_person(): + pass
\ No newline at end of file diff --git a/megapixels/commands/processor/csv_to_faces.py b/megapixels/commands/processor/csv_to_faces.py new file mode 100644 index 00000000..64c8b965 --- /dev/null +++ b/megapixels/commands/processor/csv_to_faces.py @@ -0,0 +1,105 @@ +""" +Reads in CSV of ROIs and extracts facial regions with padding +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', required=True, + help='Input CSV') +@click.option('-m', '--media', 'opt_dir_media', required=True, + help='Input image/video directory') +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory for extracted ROI images') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('--padding', 'opt_padding', default=0.25, + help='Facial padding as percentage of face width') +@click.option('--ext', 'opt_ext_out', default='png', type=click.Choice(['jpg', 'png']), + help='Output image type') +@click.option('--min', 'opt_min', default=(60, 60), + help='Minimum original face size') +@click.pass_context +def cli(ctx, opt_fp_in, opt_dir_media, opt_dir_out, opt_slice, + opt_padding, opt_ext_out, opt_min): + """Converts ROIs to images""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + from PIL import Image, ImageOps, ImageFilter, ImageDraw + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils + from app.models.bbox import BBox + + # ------------------------------------------------- + # process here + log = logger_utils.Logger.getLogger() + + df_rois = pd.read_csv(opt_fp_in, dtype={'subdir': str, 'fn': str}) + if opt_slice: + df_rois = df_rois[opt_slice[0]:opt_slice[1]] + + log.info('Processing {:,} rows'.format(len(df_rois))) + + file_utils.mkdirs(opt_dir_out) + + df_rois_grouped = df_rois.groupby(['fn']) # group by fn/filename + groups = df_rois_grouped.groups + skipped = [] + + for group in tqdm(groups): + # get image + group_rows = df_rois_grouped.get_group(group) + + row = group_rows.iloc[0] + fp_im = join(opt_dir_media, str(row['subdir']), '{fn}.{ext}'.format(**row)) # TODO change to ext + try: + im = Image.open(fp_im).convert('RGB') + im.verify() + except Exception as e: + log.warn('Could not open: {}'.format(fp_im)) + log.error(e) + continue + + for idx, roi in group_rows.iterrows(): + # get bbox to im dimensions + xywh = [roi['x'], roi['y'], roi['w'] , roi['h']] + bbox = BBox.from_xywh(*xywh) + dim = im.size + bbox_dim = bbox.to_dim(dim) + # expand + opt_padding_px = int(opt_padding * bbox_dim.width) + bbox_dim_exp = bbox_dim.expand_dim(opt_padding_px, dim) + # crop + x1y2 = bbox_dim_exp.pt_tl + bbox_dim_exp.pt_br + im_crop = im.crop(box=x1y2) + + # strip exif, create new image and paste data + im_crop_data = list(im_crop.getdata()) + im_crop_no_exif = Image.new(im_crop.mode, im_crop.size) + im_crop_no_exif.putdata(im_crop_data) + + # save + idx_zpad = file_utils.zpad(idx, zeros=3) + subdir = '' if roi['subdir'] == '.' else '{}_'.format(roi['subdir']) + subdir = subdir.replace('/', '_') + fp_im_out = join(opt_dir_out, '{}{}{}.{}'.format(subdir, roi['fn'], idx_zpad, opt_ext_out)) + # threshold size and save + if im_crop_no_exif.size[0] < opt_min[0] or im_crop_no_exif.size[1] < opt_min[1]: + skipped.append(fp_im_out) + log.info('Face too small: {}, idx: {}'.format(fp_im, idx)) + else: + im_crop_no_exif.save(fp_im_out) + + log.info('Skipped {:,} images'.format(len(skipped))) diff --git a/megapixels/commands/processor/csv_to_faces_mt.py b/megapixels/commands/processor/csv_to_faces_mt.py new file mode 100644 index 00000000..64c8b965 --- /dev/null +++ b/megapixels/commands/processor/csv_to_faces_mt.py @@ -0,0 +1,105 @@ +""" +Reads in CSV of ROIs and extracts facial regions with padding +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', required=True, + help='Input CSV') +@click.option('-m', '--media', 'opt_dir_media', required=True, + help='Input image/video directory') +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory for extracted ROI images') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('--padding', 'opt_padding', default=0.25, + help='Facial padding as percentage of face width') +@click.option('--ext', 'opt_ext_out', default='png', type=click.Choice(['jpg', 'png']), + help='Output image type') +@click.option('--min', 'opt_min', default=(60, 60), + help='Minimum original face size') +@click.pass_context +def cli(ctx, opt_fp_in, opt_dir_media, opt_dir_out, opt_slice, + opt_padding, opt_ext_out, opt_min): + """Converts ROIs to images""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + from PIL import Image, ImageOps, ImageFilter, ImageDraw + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils + from app.models.bbox import BBox + + # ------------------------------------------------- + # process here + log = logger_utils.Logger.getLogger() + + df_rois = pd.read_csv(opt_fp_in, dtype={'subdir': str, 'fn': str}) + if opt_slice: + df_rois = df_rois[opt_slice[0]:opt_slice[1]] + + log.info('Processing {:,} rows'.format(len(df_rois))) + + file_utils.mkdirs(opt_dir_out) + + df_rois_grouped = df_rois.groupby(['fn']) # group by fn/filename + groups = df_rois_grouped.groups + skipped = [] + + for group in tqdm(groups): + # get image + group_rows = df_rois_grouped.get_group(group) + + row = group_rows.iloc[0] + fp_im = join(opt_dir_media, str(row['subdir']), '{fn}.{ext}'.format(**row)) # TODO change to ext + try: + im = Image.open(fp_im).convert('RGB') + im.verify() + except Exception as e: + log.warn('Could not open: {}'.format(fp_im)) + log.error(e) + continue + + for idx, roi in group_rows.iterrows(): + # get bbox to im dimensions + xywh = [roi['x'], roi['y'], roi['w'] , roi['h']] + bbox = BBox.from_xywh(*xywh) + dim = im.size + bbox_dim = bbox.to_dim(dim) + # expand + opt_padding_px = int(opt_padding * bbox_dim.width) + bbox_dim_exp = bbox_dim.expand_dim(opt_padding_px, dim) + # crop + x1y2 = bbox_dim_exp.pt_tl + bbox_dim_exp.pt_br + im_crop = im.crop(box=x1y2) + + # strip exif, create new image and paste data + im_crop_data = list(im_crop.getdata()) + im_crop_no_exif = Image.new(im_crop.mode, im_crop.size) + im_crop_no_exif.putdata(im_crop_data) + + # save + idx_zpad = file_utils.zpad(idx, zeros=3) + subdir = '' if roi['subdir'] == '.' else '{}_'.format(roi['subdir']) + subdir = subdir.replace('/', '_') + fp_im_out = join(opt_dir_out, '{}{}{}.{}'.format(subdir, roi['fn'], idx_zpad, opt_ext_out)) + # threshold size and save + if im_crop_no_exif.size[0] < opt_min[0] or im_crop_no_exif.size[1] < opt_min[1]: + skipped.append(fp_im_out) + log.info('Face too small: {}, idx: {}'.format(fp_im, idx)) + else: + im_crop_no_exif.save(fp_im_out) + + log.info('Skipped {:,} images'.format(len(skipped))) diff --git a/megapixels/commands/processor/face_3ddfa.py b/megapixels/commands/processor/face_3ddfa.py new file mode 100644 index 00000000..ffc74180 --- /dev/null +++ b/megapixels/commands/processor/face_3ddfa.py @@ -0,0 +1,331 @@ +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, required=True, + help='Image filepath') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='GIF output path') +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('-g', '--gpu', 'opt_gpu', default=0, + help='GPU index') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('--display/--no-display', 'opt_display', is_flag=True, default=False, + help='Display detections to debug') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_gpu, opt_size, opt_force, opt_display): + """Face detector demo""" + + import sys + import os + from os.path import join + from pathlib import Path + import time + + from tqdm import tqdm + import numpy as np + import pandas as pd + import cv2 as cv + import dlib + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.utils import plot_utils + from app.processors import face_detector, face_age + from app.models.data_store import DataStore + + # 3DDFA + # git clone https://github.com/cleardusk/3DDFA/ 3rdparty/ + + import torch + import torchvision.transforms as transforms + import mobilenet_v1 + from utils.ddfa import ToTensorGjz, NormalizeGjz, str2bool + import scipy.io as sio + from utils.inference import get_suffix, parse_roi_box_from_landmark, crop_img, predict_68pts, dump_to_ply, dump_vertex, \ + draw_landmarks, predict_dense, parse_roi_box_from_bbox, get_colors, write_obj_with_colors + from utils.cv_plot import plot_pose_box + from utils.estimate_pose import parse_pose + from utils.render import get_depths_image, cget_depths_image, cpncc + from utils.paf import gen_img_paf + import argparse + import torch.backends.cudnn as cudnn + + + log = logger_utils.Logger.getLogger() + + + # ------------------------------------------------- + # load image + + im = cv.imread(opt_fp_in) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + + # ---------------------------------------------------------------------------- + # detect face + + face_detector = face_detector.DetectorDLIBCNN(gpu=opt_gpu) # -1 for CPU + bboxes = face_detector.detect(im_resized, largest=True) + bbox = bboxes[0] + dim = im_resized.shape[:2][::-1] + bbox_dim = bbox.to_dim(dim) + if not bbox: + log.error('no face detected') + return + else: + log.info(f'face detected: {bbox_dim.to_xyxy()}') + + + # ---------------------------------------------------------------------------- + # age + + age_apparent_predictor = face_age.FaceAgeApparent() + age_real_predictor = face_age.FaceAgeReal() + + st = time.time() + age_real = age_real_predictor.age(im_resized, bbox_dim) + log.info(f'age real took: {(time.time()-st)/1000:.5f}s') + st = time.time() + age_apparent = age_apparent_predictor.age(im_resized, bbox_dim) + log.info(f'age apparent took: {(time.time()-st)/1000:.5f}s') + + + # ---------------------------------------------------------------------------- + # output + + log.info(f'Face coords: {bbox_dim} face') + log.info(f'Age (real): {(age_real):.2f}') + log.info(f'Age (apparent): {(age_apparent):.2f}') + + + # ---------------------------------------------------------------------------- + # draw + + # draw real age + im_age_real = im_resized.copy() + draw_utils.draw_bbox(im_age_real, bbox_dim) + txt = f'{(age_real):.2f}' + draw_utils.draw_text(im_age_real, bbox_dim.pt_tl, txt) + + # apparent + im_age_apparent = im_resized.copy() + draw_utils.draw_bbox(im_age_apparent, bbox_dim) + txt = f'{(age_apparent):.2f}' + draw_utils.draw_text(im_age_apparent, bbox_dim.pt_tl, txt) + + + # ---------------------------------------------------------------------------- + # save + + if opt_fp_out: + # save pose only + fpp_out = Path(opt_fp_out) + + fp_out = join(fpp_out.parent, f'{fpp_out.stem}_real{fpp_out.suffix}') + cv.imwrite(fp_out, im_age_real) + + fp_out = join(fpp_out.parent, f'{fpp_out.stem}_apparent{fpp_out.suffix}') + cv.imwrite(fp_out, im_age_apparent) + + + # ---------------------------------------------------------------------------- + # display + + if opt_display: + # show all images here + cv.imshow('real', im_age_real) + cv.imshow('apparent', im_age_apparent) + display_utils.handle_keyboard() + + + + + +STD_SIZE = 120 + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='3DDFA inference pipeline') + parser.add_argument('-f', '--files', nargs='+', + help='image files paths fed into network, single or multiple images') + parser.add_argument('-m', '--mode', default='cpu', type=str, help='gpu or cpu mode') + parser.add_argument('--show_flg', default='true', type=str2bool, help='whether show the visualization result') + parser.add_argument('--bbox_init', default='one', type=str, + help='one|two: one-step bbox initialization or two-step') + parser.add_argument('--dump_res', default='true', type=str2bool, help='whether write out the visualization image') + parser.add_argument('--dump_vertex', default='true', type=str2bool, + help='whether write out the dense face vertices to mat') + parser.add_argument('--dump_ply', default='true', type=str2bool) + parser.add_argument('--dump_pts', default='true', type=str2bool) + parser.add_argument('--dump_roi_box', default='true', type=str2bool) + parser.add_argument('--dump_pose', default='true', type=str2bool) + parser.add_argument('--dump_depth', default='true', type=str2bool) + parser.add_argument('--dump_pncc', default='true', type=str2bool) + parser.add_argument('--dump_paf', default='true', type=str2bool) + parser.add_argument('--paf_size', default=3, type=int, help='PAF feature kernel size') + parser.add_argument('--dump_obj', default='true', type=str2bool) + parser.add_argument('--dlib_bbox', default='true', type=str2bool, help='whether use dlib to predict bbox') + parser.add_argument('--dlib_landmark', default='true', type=str2bool, + help='whether use dlib landmark to crop image') + + args = parser.parse_args() + main(args) + + + +def main(args): + # 1. load pre-tained model + checkpoint_fp = 'models/phase1_wpdc_vdc_v2.pth.tar' + arch = 'mobilenet_1' + + checkpoint = torch.load(checkpoint_fp, map_location=lambda storage, loc: storage)['state_dict'] + model = getattr(mobilenet_v1, arch)(num_classes=62) # 62 = 12(pose) + 40(shape) +10(expression) + model_dict = model.state_dict() + # because the model is trained by multiple gpus, prefix module should be removed + for k in checkpoint.keys(): + model_dict[k.replace('module.', '')] = checkpoint[k] + model.load_state_dict(model_dict, strict=False) + if args.mode == 'gpu': + cudnn.benchmark = True + model = model.cuda() + model.eval() + + # 2. load dlib model for face detection and landmark used for face cropping + if args.dlib_landmark: + dlib_landmark_model = 'models/shape_predictor_68_face_landmarks.dat' + face_regressor = dlib.shape_predictor(dlib_landmark_model) + if args.dlib_bbox: + face_detector = dlib.get_frontal_face_detector() + + # 3. forward + tri = sio.loadmat('visualize/tri.mat')['tri'] + transform = transforms.Compose([ToTensorGjz(), NormalizeGjz(mean=127.5, std=128)]) + for img_fp in args.files: + img_ori = cv2.imread(img_fp) + if args.dlib_bbox: + rects = face_detector(img_ori, 1) + else: + rects = [] + + if len(rects) == 0: + rects = dlib.rectangles() + rect_fp = img_fp + '.bbox' + lines = open(rect_fp).read().strip().split('\n')[1:] + for l in lines: + l, r, t, b = [int(_) for _ in l.split(' ')[1:]] + rect = dlib.rectangle(l, r, t, b) + rects.append(rect) + + pts_res = [] + Ps = [] # Camera matrix collection + poses = [] # pose collection, [todo: validate it] + vertices_lst = [] # store multiple face vertices + ind = 0 + suffix = get_suffix(img_fp) + for rect in rects: + # whether use dlib landmark to crop image, if not, use only face bbox to calc roi bbox for cropping + if args.dlib_landmark: + # - use landmark for cropping + pts = face_regressor(img_ori, rect).parts() + pts = np.array([[pt.x, pt.y] for pt in pts]).T + roi_box = parse_roi_box_from_landmark(pts) + else: + # - use detected face bbox + bbox = [rect.left(), rect.top(), rect.right(), rect.bottom()] + roi_box = parse_roi_box_from_bbox(bbox) + + img = crop_img(img_ori, roi_box) + + # forward: one step + img = cv2.resize(img, dsize=(STD_SIZE, STD_SIZE), interpolation=cv2.INTER_LINEAR) + input = transform(img).unsqueeze(0) + with torch.no_grad(): + if args.mode == 'gpu': + input = input.cuda() + param = model(input) + param = param.squeeze().cpu().numpy().flatten().astype(np.float32) + + # 68 pts + pts68 = predict_68pts(param, roi_box) + + # two-step for more accurate bbox to crop face + if args.bbox_init == 'two': + roi_box = parse_roi_box_from_landmark(pts68) + img_step2 = crop_img(img_ori, roi_box) + img_step2 = cv2.resize(img_step2, dsize=(STD_SIZE, STD_SIZE), interpolation=cv2.INTER_LINEAR) + input = transform(img_step2).unsqueeze(0) + with torch.no_grad(): + if args.mode == 'gpu': + input = input.cuda() + param = model(input) + param = param.squeeze().cpu().numpy().flatten().astype(np.float32) + + pts68 = predict_68pts(param, roi_box) + + pts_res.append(pts68) + P, pose = parse_pose(param) + Ps.append(P) + poses.append(pose) + + # dense face 3d vertices + if args.dump_ply or args.dump_vertex or args.dump_depth or args.dump_pncc or args.dump_obj: + vertices = predict_dense(param, roi_box) + vertices_lst.append(vertices) + if args.dump_ply: + dump_to_ply(vertices, tri, '{}_{}.ply'.format(img_fp.replace(suffix, ''), ind)) + if args.dump_vertex: + dump_vertex(vertices, '{}_{}.mat'.format(img_fp.replace(suffix, ''), ind)) + + # save .mat for 3d Face + wfp = '{}_{}_face3d.mat'.format(img_fp.replace(suffix, ''), ind) + colors = get_colors(img_ori, vertices) + sio.savemat(wfp, {'vertices': vertices, 'colors': colors, 'triangles': tri}) + + if args.dump_pts: + wfp = '{}_{}.txt'.format(img_fp.replace(suffix, ''), ind) + np.savetxt(wfp, pts68, fmt='%.3f') + print('Save 68 3d landmarks to {}'.format(wfp)) + if args.dump_roi_box: + wfp = '{}_{}.roibox'.format(img_fp.replace(suffix, ''), ind) + np.savetxt(wfp, roi_box, fmt='%.3f') + print('Save roi box to {}'.format(wfp)) + if args.dump_paf: + wfp_paf = '{}_{}_paf.jpg'.format(img_fp.replace(suffix, ''), ind) + wfp_crop = '{}_{}_crop.jpg'.format(img_fp.replace(suffix, ''), ind) + paf_feature = gen_img_paf(img_crop=img, param=param, kernel_size=args.paf_size) + + cv2.imwrite(wfp_paf, paf_feature) + cv2.imwrite(wfp_crop, img) + print('Dump to {} and {}'.format(wfp_crop, wfp_paf)) + if args.dump_obj: + wfp = '{}_{}.obj'.format(img_fp.replace(suffix, ''), ind) + colors = get_colors(img_ori, vertices) + write_obj_with_colors(wfp, vertices, tri, colors) + print('Dump obj with sampled texture to {}'.format(wfp)) + ind += 1 + + if args.dump_pose: + # P, pose = parse_pose(param) # Camera matrix (without scale), and pose (yaw, pitch, roll, to verify) + img_pose = plot_pose_box(img_ori, Ps, pts_res) + wfp = img_fp.replace(suffix, '_pose.jpg') + cv2.imwrite(wfp, img_pose) + print('Dump to {}'.format(wfp)) + if args.dump_depth: + wfp = img_fp.replace(suffix, '_depth.png') + # depths_img = get_depths_image(img_ori, vertices_lst, tri-1) # python version + depths_img = cget_depths_image(img_ori, vertices_lst, tri - 1) # cython version + cv2.imwrite(wfp, depths_img) + print('Dump to {}'.format(wfp)) + if args.dump_pncc: + wfp = img_fp.replace(suffix, '_pncc.png') + pncc_feature = cpncc(img_ori, vertices_lst, tri - 1) # cython version + cv2.imwrite(wfp, pncc_feature[:, :, ::-1]) # cv2.imwrite will swap RGB -> BGR + print('Dump to {}'.format(wfp)) + if args.dump_res: + draw_landmarks(img_ori, pts_res, wfp=img_fp.replace(suffix, '_3DDFA.jpg'), show_flg=args.show_flg) diff --git a/megapixels/commands/processor/face_attributes.py b/megapixels/commands/processor/face_attributes.py new file mode 100644 index 00000000..01fe3bd1 --- /dev/null +++ b/megapixels/commands/processor/face_attributes.py @@ -0,0 +1,136 @@ +""" + +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--size', 'opt_size', + type=(int, int), default=cfg.DEFAULT_SIZE_FACE_DETECT, + help='Processing size for detection') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-d', '--display', 'opt_display', is_flag=True, + help='Display image for debugging') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, + opt_size, opt_slice, opt_force, opt_display): + """Creates 2D 68-point landmarks""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.processors import face_age_gender + from app.models.data_store import DataStore + from app.models.bbox import BBox + + # ------------------------------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + # init face processors + age_estimator_apnt = face_age_gender.FaceAgeApparent() + age_estimator_real = face_age_gender.FaceAgeReal() + gender_estimator = face_age_gender.FaceGender() + + # init filepaths + data_store = DataStore(opt_data_store, opt_dataset) + # set file output path + metadata_type = types.Metadata.FACE_ATTRIBUTES + fp_out = data_store.metadata(metadata_type) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # ------------------------------------------------------------------------- + # load filepath data + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) + df_record = pd.read_csv(fp_record, dtype=cfg.FILE_RECORD_DTYPES).set_index('index') + # load ROI data + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) + df_roi = pd.read_csv(fp_roi).set_index('index') + # slice if you want + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] + # group by image index (speedup if multiple faces per image) + df_img_groups = df_roi.groupby('record_index') + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + # store landmarks in list + results = [] + + # ------------------------------------------------------------------------- + # iterate groups with file/record index as key + + for record_index, df_img_group in tqdm(df_img_groups): + + # access file_record DataSeries + file_record = df_record.iloc[record_index] + + # load image + fp_im = data_store.face(file_record.subdir, file_record.fn, file_record.ext) + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + dim = im_resized.shape[:2][::-1] + + # iterate ROIs in this image + for roi_index, df_img in df_img_group.iterrows(): + + # find landmarks + bbox_norm = BBox.from_xywh(df_img.x, df_img.y, df_img.w, df_img.h) + bbox_dim = bbox_norm.to_dim(dim) + + age_apnt = age_estimator_apnt.predict(im_resized, bbox_norm) + age_real = age_estimator_real.predict(im_resized, bbox_norm) + gender = gender_estimator.predict(im_resized, bbox_norm) + + attr_obj = { + 'age_real':float(f'{age_real:.2f}'), + 'age_apparent': float(f'{age_apnt:.2f}'), + 'm': float(f'{gender["m"]:.4f}'), + 'f': float(f'{gender["f"]:.4f}'), + 'roi_index': roi_index + } + results.append(attr_obj) + + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(results) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_frames.py b/megapixels/commands/processor/face_frames.py new file mode 100644 index 00000000..76f23af1 --- /dev/null +++ b/megapixels/commands/processor/face_frames.py @@ -0,0 +1,82 @@ +from glob import glob +import os +from os.path import join +from pathlib import Path + +import click + + + + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', required=True, + help='Input directory to glob') +@click.option('-o', '--output', 'opt_fp_out', required=True, + help='Output directory for face frames') +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_size, opt_slice): + """Split video to face frames""" + + from tqdm import tqdm + import dlib + import pandas as pd + from PIL import Image, ImageOps, ImageFilter + import cv2 as cv + import numpy as np + + from app.processors import face_detector + from app.utils import logger_utils, file_utils, im_utils + from app.settings import types + from app.utils import click_utils + from app.settings import app_cfg as cfg + from app.models.bbox import BBox + + log = logger_utils.Logger.getLogger() + + # ------------------------------------------------- + # process + + detector = face_detector.DetectorDLIBCNN() + + # get file list + fp_videos = glob(join(opt_fp_in, '*.mp4')) + fp_videos += glob(join(opt_fp_in, '*.webm')) + fp_videos += glob(join(opt_fp_in, '*.mkv')) + + min_distance_per = .025 # minimum distance percentage to save new face image + face_interval = 5 + frame_interval_count = 0 + frame_count = 0 + bbox_prev = BBox(0,0,0,0) + file_utils.mkdirs(opt_fp_out) + dnn_size = opt_size + max_dim = max(dnn_size) + px_thresh = int(max_dim * min_distance_per) + + for fp_video in tqdm(fp_videos): + # load video + video = cv.VideoCapture(fp_video) + # iterate through frames + while video.isOpened(): + res, frame = video.read() + if not res: + break + # increment frames, save frame if interval has passed + frame_count += 1 # for naming + frame_interval_count += 1 # for interval + bboxes = detector.detect(frame, opt_size=dnn_size, opt_pyramids=0) + if len(bboxes) > 0 and frame_interval_count >= face_interval: + dim = frame.shape[:2][::-1] + d = bboxes[0].to_dim(dim).distance(bbox_prev) + if d > px_thresh: + # save frame + zfc = file_utils.zpad(frame_count) + fp_frame = join(opt_fp_out, '{}_{}.jpg'.format(Path(fp_video).stem, zfc)) + cv.imwrite(fp_frame, frame) + frame_interval_count = 0 + bbox_prev = bboxes[0] diff --git a/megapixels/commands/processor/face_landmark_2d_5.py b/megapixels/commands/processor/face_landmark_2d_5.py new file mode 100644 index 00000000..40ec6f41 --- /dev/null +++ b/megapixels/commands/processor/face_landmark_2d_5.py @@ -0,0 +1,146 @@ +""" + +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +color_filters = {'color': 1, 'gray': 2, 'all': 3} + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('-d', '--detector', 'opt_detector_type', + type=cfg.FaceLandmark2D_5Var, + default=click_utils.get_default(types.FaceLandmark2D_5.DLIB), + help=click_utils.show_help(types.FaceLandmark2D_5)) +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-d', '--display', 'opt_display', is_flag=True, + help='Display image for debugging') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, opt_detector_type, + opt_size, opt_slice, opt_force, opt_display): + """Creates 2D 5-point landmarks""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.processors import face_landmarks + from app.models.data_store import DataStore + from app.models.bbox import BBox + + # ------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + # init filepaths + data_store = DataStore(opt_data_store, opt_dataset) + # set file output path + metadata_type = types.Metadata.FACE_LANDMARK_2D_5 + fp_out = data_store.metadata(metadata_type) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # init face landmark processors + if opt_detector_type == types.FaceLandmark2D_5.DLIB: + # use dlib 68 point detector + landmark_detector = face_landmarks.Dlib2D_5() + elif opt_detector_type == types.FaceLandmark2D_5.MTCNN: + # use dlib 5 point detector + landmark_detector = face_landmarks.MTCNN2D_5() + else: + log.error('{} not yet implemented'.format(opt_detector_type.name)) + return + + log.info(f'Using landmark detector: {opt_detector_type.name}') + + # load filepath data + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) + df_record = pd.read_csv(fp_record).set_index('index') + # load ROI data + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) + df_roi = pd.read_csv(fp_roi).set_index('index') + # slice if you want + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] + # group by image index (speedup if multiple faces per image) + df_img_groups = df_roi.groupby('record_index') + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + # store landmarks in list + results = [] + + # iterate groups with file/record index as key + for record_index, df_img_group in tqdm(df_img_groups): + + # acces file record + ds_record = df_record.iloc[record_index] + + # load image + fp_im = data_store.face(ds_record.subdir, ds_record.fn, ds_record.ext) + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + + # iterate image group dataframe with roi index as key + for roi_index, df_img in df_img_group.iterrows(): + + # get bbox + x, y, w, h = df_img.x, df_img.y, df_img.w, df_img.h + dim = im_resized.shape[:2][::-1] + bbox = BBox.from_xywh(x, y, w, h).to_dim(dim) + + # get landmark points + points = landmark_detector.landmarks(im_resized, bbox) + points_norm = landmark_detector.normalize(points, dim) + points_flat = landmark_detector.flatten(points_norm) + + # display to screen if optioned + if opt_display: + draw_utils.draw_landmarks2D(im_resized, points) + draw_utils.draw_bbox(im_resized, bbox) + cv.imshow('', im_resized) + display_utils.handle_keyboard() + + results.append(points_flat) + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(results) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_landmark_2d_68.py b/megapixels/commands/processor/face_landmark_2d_68.py new file mode 100644 index 00000000..c6978a40 --- /dev/null +++ b/megapixels/commands/processor/face_landmark_2d_68.py @@ -0,0 +1,150 @@ +""" + +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('-d', '--detector', 'opt_detector_type', + type=cfg.FaceLandmark2D_68Var, + default=click_utils.get_default(types.FaceLandmark2D_68.DLIB), + help=click_utils.show_help(types.FaceLandmark2D_68)) +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-d', '--display', 'opt_display', is_flag=True, + help='Display image for debugging') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, opt_detector_type, + opt_size, opt_slice, opt_force, opt_display): + """Creates 2D 68-point landmarks""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.processors import face_landmarks + from app.models.data_store import DataStore + from app.models.bbox import BBox + + # ------------------------------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + # init filepaths + data_store = DataStore(opt_data_store, opt_dataset) + # set file output path + metadata_type = types.Metadata.FACE_LANDMARK_2D_68 + fp_out = data_store.metadata(metadata_type) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # init face landmark processors + if opt_detector_type == types.FaceLandmark2D_68.DLIB: + # use dlib 68 point detector + landmark_detector = face_landmarks.Dlib2D_68() + elif opt_detector_type == types.FaceLandmark2D_68.FACE_ALIGNMENT: + # use dlib 5 point detector + landmark_detector = face_landmarks.FaceAlignment2D_68() + else: + log.error('{} not yet implemented'.format(opt_detector_type.name)) + return + + log.info(f'Using landmark detector: {opt_detector_type.name}') + + # ------------------------------------------------------------------------- + # load filepath data + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) + df_record = pd.read_csv(fp_record).set_index('index') + # load ROI data + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) + df_roi = pd.read_csv(fp_roi).set_index('index') + # slice if you want + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] + # group by image index (speedup if multiple faces per image) + df_img_groups = df_roi.groupby('record_index') + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + # store landmarks in list + results = [] + + # ------------------------------------------------------------------------- + # iterate groups with file/record index as key + + for record_index, df_img_group in tqdm(df_img_groups): + + # access file_record DataSeries + file_record = df_record.iloc[record_index] + + # load image + fp_im = data_store.face(file_record.subdir, file_record.fn, file_record.ext) + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + dim = im_resized.shape[:2][::-1] + + # iterate ROIs in this image + for roi_index, df_img in df_img_group.iterrows(): + + # find landmarks + x, y, w, h = df_img.x, df_img.y, df_img.w, df_img.h # normalized values + #dim = (file_record.width, file_record.height) # original w,h + bbox = BBox.from_xywh(x, y, w, h).to_dim(dim) + points = landmark_detector.landmarks(im_resized, bbox) + points_norm = landmark_detector.normalize(points, dim) + points_str = landmark_detector.to_str(points_norm) + + # display if optioned + if opt_display: + dst = im_resized.copy() + draw_utils.draw_landmarks2D(dst, points) + draw_utils.draw_bbox(dst, bbox) + cv.imshow('', dst) + display_utils.handle_keyboard() + + # add to results for CSV + results.append({'vec': points_str, 'roi_index':roi_index}) + + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(results) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_landmark_3d_68.py b/megapixels/commands/processor/face_landmark_3d_68.py new file mode 100644 index 00000000..a2d14d72 --- /dev/null +++ b/megapixels/commands/processor/face_landmark_3d_68.py @@ -0,0 +1,147 @@ +""" + +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +color_filters = {'color': 1, 'gray': 2, 'all': 3} + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('-d', '--detector', 'opt_detector_type', + type=cfg.FaceLandmark3D_68Var, + default=click_utils.get_default(types.FaceLandmark3D_68.FACE_ALIGNMENT), + help=click_utils.show_help(types.FaceLandmark3D_68)) +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-d', '--display', 'opt_display', is_flag=True, + help='Display image for debugging') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, opt_detector_type, + opt_size, opt_slice, opt_force, opt_display): + """Generate 3D 68-point landmarks""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.utils import plot_utils + from app.processors import face_landmarks + from app.models.data_store import DataStore + from app.models.bbox import BBox + + # -------------------------------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + log.warn('not normalizing points') + # init filepaths + data_store = DataStore(opt_data_store, opt_dataset) + # set file output path + metadata_type = types.Metadata.FACE_LANDMARK_3D_68 + fp_out = data_store.metadata(metadata_type) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # init face landmark processors + if opt_detector_type == types.FaceLandmark3D_68.FACE_ALIGNMENT: + # use FaceAlignment 68 point 3D detector + landmark_detector = face_landmarks.FaceAlignment3D_68() + else: + log.error('{} not yet implemented'.format(opt_detector_type.name)) + return + + log.info(f'Using landmark detector: {opt_detector_type.name}') + + # ------------------------------------------------------------------------- + # load data + + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) # file_record.csv + df_record = pd.read_csv(fp_record).set_index('index') + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) # face_roi.csv + df_roi = pd.read_csv(fp_roi).set_index('index') + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] # slice if you want + df_img_groups = df_roi.groupby('record_index') # groups by image index (load once) + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + # store landmarks in list + results = [] + + # iterate groups with file/record index as key + for record_index, df_img_group in tqdm(df_img_groups): + + # acces file record + ds_record = df_record.iloc[record_index] + + # load image + fp_im = data_store.face(ds_record.subdir, ds_record.fn, ds_record.ext) + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + + # iterate image group dataframe with roi index as key + for roi_index, df_img in df_img_group.iterrows(): + + # get bbox + x, y, w, h = df_img.x, df_img.y, df_img.w, df_img.h + dim = im_resized.shape[:2][::-1] + bbox = BBox.from_xywh(x, y, w, h).to_dim(dim) + + # get landmark points + points = landmark_detector.landmarks(im_resized, bbox) + # NB can't really normalize these points, but are normalized against 3D space + #points_norm = landmark_detector.normalize(points, dim) # normalized using 200 + points_flattenend = landmark_detector.flatten(points) + + # display to screen if optioned + if opt_display: + draw_utils.draw_landmarks3D(im_resized, points) + draw_utils.draw_bbox(im_resized, bbox) + cv.imshow('', im_resized) + display_utils.handle_keyboard() + + #plot_utils.generate_3d_landmark_anim(points, '/home/adam/Downloads/3d.gif') + + results.append(points_flattenend) + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(results) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_pose.py b/megapixels/commands/processor/face_pose.py new file mode 100644 index 00000000..cb7ec56c --- /dev/null +++ b/megapixels/commands/processor/face_pose.py @@ -0,0 +1,164 @@ +""" +NB: This only works with the DLIB 68-point landmarks. + +Converts ROIs to pose: yaw, roll, pitch +pitch: looking down or up in yes gesture +roll: tilting head towards shoulder +yaw: twisting head left to right in no gesture + +""" + +""" +TODO +- check compatibility with MTCNN 68 point detector +- improve accuracy by using MTCNN 5-point +- refer to https://github.com/jerryhouuu/Face-Yaw-Roll-Pitch-from-Pose-Estimation-using-OpenCV/ +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--size', 'opt_size', + type=(int, int), default=(300, 300), + help='Output image size') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-d', '--display', 'opt_display', is_flag=True, + help='Display image for debugging') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, opt_size, + opt_slice, opt_force, opt_display): + """Converts ROIs to pose: roll, yaw, pitch""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import dlib # must keep a local reference for dlib + import cv2 as cv + import pandas as pd + + from app.models.bbox import BBox + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.processors.face_landmarks import Dlib2D_68 + from app.processors.face_pose import FacePoseDLIB + from app.models.data_store import DataStore + + # ------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + + # set data_store + data_store = DataStore(opt_data_store, opt_dataset) + + # get filepath out + fp_out = data_store.metadata(types.Metadata.FACE_POSE) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # init face processors + face_pose = FacePoseDLIB() + face_landmarks = Dlib2D_68() + + # ------------------------------------------------- + # load data + + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) + df_record = pd.read_csv(fp_record, dtype=cfg.FILE_RECORD_DTYPES).set_index('index') + # load ROI data + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) + df_roi = pd.read_csv(fp_roi).set_index('index') + # slice if you want + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] + # group by image index (speedup if multiple faces per image) + df_img_groups = df_roi.groupby('record_index') + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + # store poses and convert to DataFrame + results = [] + + # ------------------------------------------------- + # iterate groups with file/record index as key + for record_index, df_img_group in tqdm(df_img_groups): + + # access the file_record + file_record = df_record.iloc[record_index] # pands.DataSeries + + # load image + fp_im = data_store.face(file_record.subdir, file_record.fn, file_record.ext) + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + + # iterate image group dataframe with roi index as key + for roi_index, df_img in df_img_group.iterrows(): + + # get bbox + x, y, w, h = df_img.x, df_img.y, df_img.w, df_img.h + #dim = (file_record.width, file_record.height) + dim = im_resized.shape[:2][::-1] + bbox_norm = BBox.from_xywh(x, y, w, h) + bbox_dim = bbox_norm.to_dim(dim) + + # get pose + landmarks = face_landmarks.landmarks(im_resized, bbox_norm) + pose_data = face_pose.pose(landmarks, dim) + #pose_degrees = pose_data['degrees'] # only keep the degrees data + #pose_degrees['points_nose'] = pose_data + + # draw landmarks if optioned + if opt_display: + draw_utils.draw_pose(im_resized, pose_data['point_nose'], pose_data['points']) + draw_utils.draw_degrees(im_resized, pose_data) + cv.imshow('', im_resized) + display_utils.handle_keyboard() + + # add image index and append to result CSV data + pose_data['roi_index'] = roi_index + for k, v in pose_data['points'].items(): + pose_data[f'point_{k}_x'] = v[0] / dim[0] + pose_data[f'point_{k}_y'] = v[1] / dim[1] + + # rearrange data structure for DataFrame + pose_data.pop('points') + pose_data['point_nose_x'] = pose_data['point_nose'][0] / dim[0] + pose_data['point_nose_y'] = pose_data['point_nose'][1] / dim[1] + pose_data.pop('point_nose') + results.append(pose_data) + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(results) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_roi.py b/megapixels/commands/processor/face_roi.py new file mode 100644 index 00000000..fc933049 --- /dev/null +++ b/megapixels/commands/processor/face_roi.py @@ -0,0 +1,187 @@ +""" +Crop images to prepare for training +""" + +import click +# from PIL import Image, ImageOps, ImageFilter, ImageDraw + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +color_filters = {'color': 1, 'gray': 2, 'all': 3} + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', default=None, + help='Override enum input filename CSV') +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--size', 'opt_size', + type=(int, int), default=(480, 480), + help='Output image size') +@click.option('-d', '--detector', 'opt_detector_type', + type=cfg.FaceDetectNetVar, + default=click_utils.get_default(types.FaceDetectNet.CVDNN), + help=click_utils.show_help(types.FaceDetectNet)) +@click.option('-g', '--gpu', 'opt_gpu', default=0, + help='GPU index') +@click.option('--conf', 'opt_conf_thresh', default=0.85, type=click.FloatRange(0,1), + help='Confidence minimum threshold') +@click.option('-p', '--pyramids', 'opt_pyramids', default=0, type=click.IntRange(0,4), + help='Number pyramids to upscale for DLIB detectors') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('--display/--no-display', 'opt_display', is_flag=True, default=False, + help='Display detections to debug') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('--color', 'opt_color_filter', + type=click.Choice(color_filters.keys()), default='color', + help='Filter to keep color or grayscale images (color = keep color') +@click.option('--keep', 'opt_largest', type=click.Choice(['largest', 'all']), default='largest', + help='Only keep largest face') +@click.option('--zone', 'opt_zone', default=(0.0, 0.0), type=(float, float), + help='Face center must be located within zone region (0.5 = half width/height)') +@click.pass_context +def cli(ctx, opt_fp_in, opt_dir_media, opt_fp_out, opt_data_store, opt_dataset, opt_size, opt_detector_type, + opt_gpu, opt_conf_thresh, opt_pyramids, opt_slice, opt_display, opt_force, opt_color_filter, + opt_largest, opt_zone): + """Converts frames with faces to CSV of ROIs""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import dlib # must keep a local reference for dlib + import cv2 as cv + import pandas as pd + + from app.utils import logger_utils, file_utils, im_utils, display_utils, draw_utils + from app.processors import face_detector + from app.models.data_store import DataStore + + # ------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + + # set data_store + data_store = DataStore(opt_data_store, opt_dataset) + + # get filepath out + fp_out = data_store.metadata(types.Metadata.FACE_ROI) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # set detector + if opt_detector_type == types.FaceDetectNet.CVDNN: + detector = face_detector.DetectorCVDNN() + elif opt_detector_type == types.FaceDetectNet.DLIB_CNN: + detector = face_detector.DetectorDLIBCNN(gpu=opt_gpu) + elif opt_detector_type == types.FaceDetectNet.DLIB_HOG: + detector = face_detector.DetectorDLIBHOG() + elif opt_detector_type == types.FaceDetectNet.MTCNN_TF: + detector = face_detector.DetectorMTCNN_TF(gpu=opt_gpu) + elif opt_detector_type == types.FaceDetectNet.HAAR: + log.error('{} not yet implemented'.format(opt_detector_type.name)) + return + + + # get list of files to process + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) if opt_fp_in is None else opt_fp_in + df_record = pd.read_csv(fp_record, dtype=cfg.FILE_RECORD_DTYPES).set_index('index') + if opt_slice: + df_record = df_record[opt_slice[0]:opt_slice[1]] + log.debug('processing {:,} files'.format(len(df_record))) + + # filter out grayscale + color_filter = color_filters[opt_color_filter] + # set largest flag, to keep all or only largest + opt_largest = (opt_largest == 'largest') + + data = [] + skipped_files = [] + processed_files = [] + + for df_record in tqdm(df_record.itertuples(), total=len(df_record)): + fp_im = data_store.face(str(df_record.subdir), str(df_record.fn), str(df_record.ext)) + try: + im = cv.imread(fp_im) + im_resized = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + except Exception as e: + log.debug(f'could not read: {fp_im}') + return + # filter out color or grayscale iamges + if color_filter != color_filters['all']: + try: + is_gray = im_utils.is_grayscale(im) + if is_gray and color_filter != color_filters['gray']: + log.debug('Skipping grayscale image: {}'.format(fp_im)) + continue + except Exception as e: + log.error('Could not check grayscale: {}'.format(fp_im)) + continue + + try: + bboxes_norm = detector.detect(im_resized, pyramids=opt_pyramids, largest=opt_largest, + zone=opt_zone, conf_thresh=opt_conf_thresh) + except Exception as e: + log.error('could not detect: {}'.format(fp_im)) + log.error('{}'.format(e)) + continue + + if len(bboxes_norm) == 0: + skipped_files.append(fp_im) + log.warn(f'no faces in: {fp_im}') + log.warn(f'skipped: {len(skipped_files)}. found:{len(processed_files)} files') + else: + processed_files.append(fp_im) + for bbox in bboxes_norm: + roi = { + 'record_index': int(df_record.Index), + 'x': bbox.x, + 'y': bbox.y, + 'w': bbox.w, + 'h': bbox.h + } + data.append(roi) + + # if display optined + if opt_display and len(bboxes_norm): + # draw each box + for bbox_norm in bboxes_norm: + dim = im_resized.shape[:2][::-1] + bbox_dim = bbox.to_dim(dim) + if dim[0] > 1000: + im_resized = im_utils.resize(im_resized, width=1000) + im_resized = draw_utils.draw_bbox(im_resized, bbox_norm) + + # display and wait + cv.imshow('', im_resized) + display_utils.handle_keyboard() + + # create DataFrame and save to CSV + file_utils.mkdirs(fp_out) + df = pd.DataFrame.from_dict(data) + df.index.name = 'index' + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/face_vector.py b/megapixels/commands/processor/face_vector.py new file mode 100644 index 00000000..cb155d08 --- /dev/null +++ b/megapixels/commands/processor/face_vector.py @@ -0,0 +1,133 @@ +""" +Converts ROIs to face vector +NB: the VGG Face2 extractor should be used with MTCNN ROIs (not square) + the DLIB face extractor should be used with DLIB ROIs (square) +see https://github.com/ox-vgg/vgg_face2 for TAR@FAR +""" + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +@click.command() +@click.option('-o', '--output', 'opt_fp_out', default=None, + help='Override enum output filename CSV') +@click.option('-m', '--media', 'opt_dir_media', default=None, + help='Override enum media directory') +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--size', 'opt_size', + type=(int, int), default=cfg.DEFAULT_SIZE_FACE_DETECT, + help='Output image size') +@click.option('-e', '--extractor', 'opt_extractor', + default=click_utils.get_default(types.FaceExtractor.VGG), + type=cfg.FaceExtractorVar, + help='Type of extractor framework/network to use') +@click.option('-j', '--jitters', 'opt_jitters', default=cfg.DLIB_FACEREC_JITTERS, + help='Number of jitters (only for dlib') +@click.option('-p', '--padding', 'opt_padding', default=cfg.FACEREC_PADDING, + help='Percentage ROI padding') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice list of files') +@click.option('-f', '--force', 'opt_force', is_flag=True, + help='Force overwrite file') +@click.option('-g', '--gpu', 'opt_gpu', default=0, + help='GPU index') +@click.pass_context +def cli(ctx, opt_fp_out, opt_dir_media, opt_data_store, opt_dataset, opt_size, + opt_extractor, opt_slice, opt_force, opt_gpu, opt_jitters, opt_padding): + """Converts face ROIs to vectors""" + + import sys + import os + from os.path import join + from pathlib import Path + from glob import glob + + from tqdm import tqdm + import numpy as np + import dlib # must keep a local reference for dlib + import cv2 as cv + import pandas as pd + + from app.models.bbox import BBox + from app.models.data_store import DataStore + from app.utils import logger_utils, file_utils, im_utils + from app.processors import face_extractor + + + # ------------------------------------------------- + # init here + + log = logger_utils.Logger.getLogger() + # set data_store + data_store = DataStore(opt_data_store, opt_dataset) + + # get filepath out + fp_out = data_store.metadata(types.Metadata.FACE_VECTOR) if opt_fp_out is None else opt_fp_out + if not opt_force and Path(fp_out).exists(): + log.error('File exists. Use "-f / --force" to overwite') + return + + # init face processors + if opt_extractor == types.FaceExtractor.DLIB: + log.debug('set dlib') + extractor = face_extractor.ExtractorDLIB(gpu=opt_gpu, jitters=opt_jitters) + elif opt_extractor == types.FaceExtractor.VGG: + extractor = face_extractor.ExtractorVGG() + + # load data + fp_record = data_store.metadata(types.Metadata.FILE_RECORD) + df_record = pd.read_csv(fp_record, dtype=cfg.FILE_RECORD_DTYPES).set_index('index') + fp_roi = data_store.metadata(types.Metadata.FACE_ROI) + df_roi = pd.read_csv(fp_roi).set_index('index') + + if opt_slice: + df_roi = df_roi[opt_slice[0]:opt_slice[1]] + + # ------------------------------------------------- + # process images + + df_img_groups = df_roi.groupby('record_index') + log.debug('processing {:,} groups'.format(len(df_img_groups))) + + vecs = [] + for record_index, df_img_group in tqdm(df_img_groups): + # make fp + ds_record = df_record.iloc[record_index] + fp_im = data_store.face(ds_record.subdir, ds_record.fn, ds_record.ext) + im = cv.imread(fp_im) + im = im_utils.resize(im, width=opt_size[0], height=opt_size[1]) + for roi_index, df_img in df_img_group.iterrows(): + # get bbox + x, y, w, h = df_img.x, df_img.y, df_img.w, df_img.h + dim = (ds_record.width, ds_record.height) + # get face vector + bbox = BBox.from_xywh(x, y, w, h) # norm + # compute vec + vec = extractor.extract(im, bbox) # use normalized BBox + vec_str = extractor.to_str(vec) + vec_obj = {'vec':vec_str, 'roi_index': roi_index, 'record_index':record_index} + vecs.append(vec_obj) + + # ------------------------------------------------- + # save data + + # create DataFrame and save to CSV + df = pd.DataFrame.from_dict(vecs) + df.index.name = 'index' + file_utils.mkdirs(fp_out) + df.to_csv(fp_out) + + # save script + file_utils.write_text(' '.join(sys.argv), '{}.sh'.format(fp_out))
\ No newline at end of file diff --git a/megapixels/commands/processor/mirror.py b/megapixels/commands/processor/mirror.py new file mode 100644 index 00000000..9ca1cac7 --- /dev/null +++ b/megapixels/commands/processor/mirror.py @@ -0,0 +1,57 @@ +""" +Crop images to prepare for training +""" + +import click +import cv2 as cv +from PIL import Image, ImageOps, ImageFilter + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + + +@click.command() +@click.option('-i', '--input', 'opt_dir_in', required=True, + help='Input directory') +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice the input list') +@click.pass_context +def cli(ctx, opt_dir_in, opt_dir_out, opt_slice): + """Mirror augment image directory""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + from tqdm import tqdm + + from app.utils import logger_utils, file_utils, im_utils + + # ------------------------------------------------- + # init + + log = logger_utils.Logger.getLogger() + + # ------------------------------------------------- + # process here + + # get list of files to process + fp_ims = glob(join(opt_dir_in, '*.jpg')) + fp_ims += glob(join(opt_dir_in, '*.png')) + + if opt_slice: + fp_ims = fp_ims[opt_slice[0]:opt_slice[1]] + log.info('processing {:,} files'.format(len(fp_ims))) + + # ensure output dir exists + file_utils.mkdirs(opt_dir_out) + + # resize and save images + for fp_im in tqdm(fp_ims): + im = Image.open(fp_im) + fpp_im = Path(fp_im) + fp_out = join(opt_dir_out, '{}_mirror{}'.format(fpp_im.stem, fpp_im.suffix)) + im.save(fp_out)
\ No newline at end of file diff --git a/megapixels/commands/processor/resize.py b/megapixels/commands/processor/resize.py new file mode 100644 index 00000000..7409ee6f --- /dev/null +++ b/megapixels/commands/processor/resize.py @@ -0,0 +1,150 @@ +""" +Crop images to prepare for training +""" + +import click +import cv2 as cv +from PIL import Image, ImageOps, ImageFilter + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +""" +Filter Q-Down Q-Up Speed +NEAREST ⭐⭐⭐⭐⭐ +BOX ⭐ ⭐⭐⭐⭐ +BILINEAR ⭐ ⭐ ⭐⭐⭐ +HAMMING ⭐⭐ ⭐⭐⭐ +BICUBIC ⭐⭐⭐ ⭐⭐⭐ ⭐⭐ +LANCZOS ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐ +""" +methods = { + 'lanczos': Image.LANCZOS, + 'bicubic': Image.BICUBIC, + 'hamming': Image.HAMMING, + 'bileaner': Image.BILINEAR, + 'box': Image.BOX, + 'nearest': Image.NEAREST + } +centerings = { + 'tl': (0.0, 0.0), + 'tc': (0.5, 0.0), + 'tr': (0.0, 0.0), + 'lc': (0.0, 0.5), + 'cc': (0.5, 0.5), + 'rc': (1.0, 0.5), + 'bl': (0.0, 1.0), + 'bc': (1.0, 0.5), + 'br': (1.0, 1.0) +} + +@click.command() +@click.option('-i', '--input', 'opt_dir_in', required=True, + help='Input directory') +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory') +@click.option('-e', '--ext', 'opt_glob_ext', + default='png', type=click.Choice(['jpg', 'png']), + help='File glob ext') +@click.option('--size', 'opt_size', + type=(int, int), default=(256, 256), + help='Max output size') +@click.option('--method', 'opt_scale_method', + type=click.Choice(methods.keys()), + default='lanczos', + help='Scaling method to use') +@click.option('--equalize', 'opt_equalize', is_flag=True, + help='Equalize historgram') +@click.option('--sharpen', 'opt_sharpen', is_flag=True, + help='Unsharp mask') +@click.option('--center', 'opt_center', default='cc', type=click.Choice(centerings.keys()), + help='Crop focal point') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice the input list') +@click.option('-t', '--threads', 'opt_threads', default=8, + help='Number of threads') +@click.pass_context +def cli(ctx, opt_dir_in, opt_dir_out, opt_glob_ext, opt_size, opt_scale_method, + opt_equalize, opt_sharpen, opt_center, opt_slice, opt_threads): + """Crop, mirror images""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + from tqdm import tqdm + from multiprocessing.dummy import Pool as ThreadPool + from functools import partial + + from app.utils import logger_utils, file_utils, im_utils + + # ------------------------------------------------- + # init + + log = logger_utils.Logger.getLogger() + + + # ------------------------------------------------- + # process here + + def pool_resize(fp_im, opt_size, scale_method): + # Threaded image resize function + try: + pbar.update(1) + try: + im = Image.open(fp_im).convert('RGB') + im.verify() + except Exception as e: + log.warn('Could not open: {}'.format(fp_im)) + log.error(e) + return False + + #im = ImageOps.fit(im, opt_size, method=scale_method, centering=centering) + + if opt_equalize: + im_np = im_utils.pil2np(im) + im_np_eq = eq_hist_yuv(im_np) + im_np = cv.addWeighted(im_np_eq, 0.35, im_np, 0.65, 0) + im = im_utils.np2pil(im_np) + + if opt_sharpen: + im = im.filter(ImageFilter.UnsharpMask) + + fp_out = join(opt_dir_out, Path(fp_im).name) + im.save(fp_out) + return True + except: + return False + + #centering = centerings[opt_center] + #scale_method = methods[opt_scale_method] + + # get list of files to process + fp_ims = glob(join(opt_dir_in, '*.{}'.format(opt_glob_ext))) + if opt_slice: + fp_ims = fp_ims[opt_slice[0]:opt_slice[1]] + log.info('processing {:,} files'.format(len(fp_ims))) + + + # ensure output dir exists + file_utils.mkdirs(opt_dir_out) + + # setup multithreading + pbar = tqdm(total=len(fp_ims)) + #pool_resize = partial(pool_resize, opt_size=opt_size, scale_method=scale_method, centering=centering) + pool_resize = partial(pool_resize, opt_size=opt_size) + #result_list = pool.map(prod_x, data_list) + pool = ThreadPool(opt_threads) + with tqdm(total=len(fp_ims)) as pbar: + results = pool.map(pool_resize, fp_ims) + pbar.close() + + log.info('Resized: {} / {} images'.format(results.count(True), len(fp_ims))) + + + +def eq_hist_yuv(im): + im_yuv = cv.cvtColor(im, cv.COLOR_BGR2YUV) + im_yuv[:,:,0] = cv.equalizeHist(im_yuv[:,:,0]) + return cv.cvtColor(im_yuv, cv.COLOR_YUV2BGR) diff --git a/megapixels/commands/processor/resize_dataset.py b/megapixels/commands/processor/resize_dataset.py new file mode 100644 index 00000000..3a6ec15f --- /dev/null +++ b/megapixels/commands/processor/resize_dataset.py @@ -0,0 +1,149 @@ +""" +Crop images to prepare for training +""" + +import click +import cv2 as cv +from PIL import Image, ImageOps, ImageFilter + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg + +cv_resize_algos = { + 'area': cv.INTER_AREA, + 'lanco': cv.INTER_LANCZOS4, + 'linear': cv.INTER_LINEAR, + 'linear_exact': cv.INTER_LINEAR_EXACT, + 'nearest': cv.INTER_NEAREST +} +""" +Filter Q-Down Q-Up Speed +NEAREST ⭐⭐⭐⭐⭐ +BOX ⭐ ⭐⭐⭐⭐ +BILINEAR ⭐ ⭐ ⭐⭐⭐ +HAMMING ⭐⭐ ⭐⭐⭐ +BICUBIC ⭐⭐⭐ ⭐⭐⭐ ⭐⭐ +LANCZOS ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐ +""" +pil_resize_algos = { + 'antialias': Image.ANTIALIAS, + 'lanczos': Image.LANCZOS, + 'bicubic': Image.BICUBIC, + 'hamming': Image.HAMMING, + 'bileaner': Image.BILINEAR, + 'box': Image.BOX, + 'nearest': Image.NEAREST + } + +@click.command() +@click.option('--dataset', 'opt_dataset', + type=cfg.DatasetVar, + required=True, + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('--store', 'opt_data_store', + type=cfg.DataStoreVar, + default=click_utils.get_default(types.DataStore.HDD), + show_default=True, + help=click_utils.show_help(types.Dataset)) +@click.option('-o', '--output', 'opt_dir_out', required=True, + help='Output directory') +@click.option('-e', '--ext', 'opt_glob_ext', + default='png', type=click.Choice(['jpg', 'png']), + help='File glob ext') +@click.option('--size', 'opt_size', + type=(int, int), default=(256, 256), + help='Output image size max (w,h)') +@click.option('--interp', 'opt_interp_algo', + type=click.Choice(pil_resize_algos.keys()), + default='bicubic', + help='Interpolation resizing algorithms') +@click.option('--slice', 'opt_slice', type=(int, int), default=(None, None), + help='Slice the input list') +@click.option('-t', '--threads', 'opt_threads', default=8, + help='Number of threads') +@click.option('--recursive/--no-recursive', 'opt_recursive', is_flag=True, default=False, + help='Use glob recursion (slower)') +@click.pass_context +def cli(ctx, opt_dataset, opt_data_store, opt_dir_out, opt_glob_ext, opt_size, opt_interp_algo, + opt_slice, opt_threads, opt_recursive): + """Resize dataset images""" + + import os + from os.path import join + from pathlib import Path + from glob import glob + from tqdm import tqdm + from multiprocessing.dummy import Pool as ThreadPool + from functools import partial + import pandas as pd + import numpy as np + + from app.utils import logger_utils, file_utils, im_utils + from app.models.data_store import DataStore + + # ------------------------------------------------- + # init + + log = logger_utils.Logger.getLogger() + + + # ------------------------------------------------- + # process here + + def pool_resize(fp_in, dir_in, dir_out, im_size, interp_algo): + # Threaded image resize function + pbar.update(1) + try: + im = Image.open(fp_in).convert('RGB') + im.verify() # throws error if image is corrupt + im.thumbnail(im_size, interp_algo) + fp_out = fp_in.replace(dir_in, dir_out) + file_utils.mkdirs(fp_out) + im.save(fp_out, quality=100) + except Exception as e: + log.warn(f'Could not open: {fp_in}, Error: {e}') + return False + return True + + + data_store = DataStore(opt_data_store, opt_dataset) + fp_records = data_store.metadata(types.Metadata.FILE_RECORD) + df_records = pd.read_csv(fp_records, dtype=cfg.FILE_RECORD_DTYPES).set_index('index') + dir_in = data_store.media_images_original() + + # get list of files to process + #fp_ims = file_utils.glob_multi(opt_dir_in, ['jpg', 'png'], recursive=opt_recursive) + fp_ims = [] + for ds_record in df_records.itertuples(): + fp_im = data_store.face(ds_record.subdir, ds_record.fn, ds_record.ext) + fp_ims.append(fp_im) + + if opt_slice: + fp_ims = fp_ims[opt_slice[0]:opt_slice[1]] + if not fp_ims: + log.error('No images. Try with "--recursive"') + return + log.info(f'processing {len(fp_ims):,} images') + + # algorithm to use for resizing + interp_algo = pil_resize_algos[opt_interp_algo] + log.info(f'using {interp_algo} for interpoloation') + + # ensure output dir exists + file_utils.mkdirs(opt_dir_out) + + # setup multithreading + pbar = tqdm(total=len(fp_ims)) + # fixed arguments for pool function + map_pool_resize = partial(pool_resize, dir_in=dir_in, dir_out=opt_dir_out, im_size=opt_size, interp_algo=interp_algo) + #result_list = pool.map(prod_x, data_list) # simple + pool = ThreadPool(opt_threads) + # start multithreading + with tqdm(total=len(fp_ims)) as pbar: + results = pool.map(map_pool_resize, fp_ims) + # end multithreading + pbar.close() + + log.info(f'Resized: {results.count(True)} / {len(fp_ims)} images')
\ No newline at end of file diff --git a/megapixels/commands/processor/videos_to_frames.py b/megapixels/commands/processor/videos_to_frames.py new file mode 100644 index 00000000..0b56c46a --- /dev/null +++ b/megapixels/commands/processor/videos_to_frames.py @@ -0,0 +1,73 @@ +from glob import glob +import os +from os.path import join +from pathlib import Path + +import click + +from app.settings import types +from app.utils import click_utils +from app.settings import app_cfg as cfg +from app.utils import logger_utils + +import dlib +import pandas as pd +from PIL import Image, ImageOps, ImageFilter +from app.utils import file_utils, im_utils + + +log = logger_utils.Logger.getLogger() + +@click.command() +@click.option('-i', '--input', 'opt_fp_in', required=True, + help='Input directory') +@click.option('-o', '--output', 'opt_fp_out', required=True, + help='Output directory') +@click.option('--size', 'opt_size', default=(320, 240), + help='Inference size for face detection' ) +@click.option('--interval', 'opt_frame_interval', default=20, + help='Number of frames before saving next face') +@click.pass_context +def cli(ctx, opt_fp_in, opt_fp_out, opt_size, opt_frame_interval): + """Converts videos to frames with faces""" + + # ------------------------------------------------- + # process + + from tqdm import tqdm + import cv2 as cv + from tqdm import tqdm + from app.processors import face_detector + + detector = face_detector.DetectorDLIBCNN() + + # get file list + fp_videos = glob(join(opt_fp_in, '*.mp4')) + fp_videos += glob(join(opt_fp_in, '*.webm')) + fp_videos += glob(join(opt_fp_in, '*.mkv')) + + frame_interval_count = 0 + frame_count = 0 + + file_utils.mkdirs(opt_fp_out) + + for fp_video in tqdm(fp_videos): + + video = cv.VideoCapture(fp_video) + + while video.isOpened(): + res, frame = video.read() + if not res: + break + + frame_count += 1 # for naming + frame_interval_count += 1 # for interval + + bboxes = detector.detect(frame, opt_size=opt_size, opt_pyramids=0) + if len(bboxes) > 0 and frame_interval_count >= opt_frame_interval: + # save frame + fname = file_utils.zpad(frame_count) + fp_frame = join(opt_fp_out, '{}_{}.jpg'.format(Path(fp_video).stem, fname)) + cv.imwrite(fp_frame, frame) + frame_interval_count = 0 + |