def preprocess(tensor):
g = np.random.randint(0, 7)
n = np.random.sample() * 0.01
tensor = tensor + n
if g > 0:
degrade_blur = GaussianBlur_CUDA(g / 18)
tensor = degrade_blur(tensor)
pre_blur = GaussianBlur_CUDA(1)
tensor = pre_blur(tensor)
if np.random.sample() < 0.5:
ww = rand_motion_blur_weight(3, 5, 360).cuda()
tensor = torch.nn.functional.conv2d(tensor, ww, None, 1, ww.shape[2] // 2)
return tensor
def main(args):
# 1. load pre-tained model
checkpoint_fp = 'models/phase1_wpdc_vdc.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)
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)])
filenames = [join(args.files, x) for x in listdir(args.files) if is_image_file(x)]
filenames.sort()
result_fn = join(args.files, 'results')
if not os.path.exists(result_fn):
os.mkdir(result_fn)
blur = GaussianBlur_CUDA(sigma=35)
rgb2y = RGB2Y('cuda')
for img_fp in filenames:
print("process ", img_fp)
img_ori = cv2.imread(img_fp)
img_data = torch.Tensor(cv2.cvtColor(img_ori, cv2.COLOR_BGR2RGB)).permute(2, 0, 1) / 255.0
# print(img_data.shape)
Y = blur(rgb2y.do(img_data.to('cuda').unsqueeze(0)))
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)
# print(pts68, len(pts68))
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 or args.dump_res:
vertices = predict_dense(param, roi_box)
# print('vertex size is = ', len(vertices), vertices.shape, "param len ", len(param))
vertices_lst.append(vertices)
# print(param)
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))
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:
# print(pts_res, len(pts_res), len(pts_res[0][0]))
# draw_landmarks(img_ori, pts_res, wfp=img_fp.replace(suffix, '_3DDFA.jpg'), show_flg=args.show_flg, color='yellow')
# print("dump res")
h, w = img_ori.shape[:2]
blank = np.zeros((h, w, 3))
result = plot_kpt(blank, pts_res)
# print(result.shape)
bn = basename(img_fp)
ofn = join(result_fn, bn).replace(suffix, '_landmark.jpg')
cv2.imwrite(ofn, result)
# tri2 = tri / 29
# print(tri2, tri)
result = plot_mesh_simple(blank, vertices_lst)
ofn = join(result_fn, bn).replace(suffix, '_mesh.jpg')
cv2.imwrite(ofn, result)
result = plot_light(blank, vertices_lst, Y)
ofn = join(result_fn, bn).replace(suffix, '_light.jpg')
cv2.imwrite(ofn, result)
# print("input size", ow, oh, sz)
pad = []
pad.append((sz - ow) // 2)
pad.append(sz - ow - pad[0])
pad.append((sz - oh) // 2)
pad.append(sz - oh - pad[2])
t = torch.nn.functional.pad(t, pad, mode='constant', value=0.5)
t = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])(t)
t = t.cuda().unsqueeze(0)
t = torch.nn.functional.interpolate(t, (256, 256), mode='bilinear', align_corners=False)
blur = GaussianBlur_CUDA(0.5)
# t = blur(t)
t1 = t.clone()
t2 = t.clone()
outputs = net(t, t1, t2)
heatmaps = outputs[-2][:, :17] # .squeeze(0)
paf_maps = outputs[-1] # .squeeze(0)
# heatmap = heatmaps.sum(1, keepdim=True)
# heatmap = heatmap[:, :, :255].sum(2)
heatmap = heatmaps.squeeze(0).reshape(17, 1, heatmaps.shape[2], heatmaps.shape[3])
viz.draw_images(heatmap, "output_heatmap")
# opt.face_cutter = FaceCutter(512,
# cuda=True,
# blur_threshold=opt.config.dface.blur_threshold,
# method=opt.config.dface.method,
# mtcnn_path=opt.config.dface.mtcnn_path,
# small_face_size=opt.config.dface.small_face_size,
# tiny_face_size=opt.config.dface.tiny_face_size,
# mtcnn_thresholds=opt.config.dface.mtcnn_thresholds,
# mtcnn_nms_thresholds=opt.config.dface.mtcnn_nms_thresholds
# )
from DGPT.Utils.CUDAFuncs.GaussianBlur import GaussianBlur_CUDA
# opt.face_initial_blur = GaussianBlur_CUDA(opt.face_net.init_sigma)
print("base sigma=", opt.config.face_net.base_sigma)
opt.face_base_blur = GaussianBlur_CUDA(opt.config.face_net.base_sigma)
opt.face_valid_blur = GaussianBlur_CUDA(opt.config.face_net.valid_sigma)
opt.bg_valid_blur = GaussianBlur_CUDA(opt.config.bg_net.valid_sigma)
opt.bg_base_blur = GaussianBlur_CUDA(opt.config.bg_net.base_sigma)
# from quicklab.utils import VggUnNormalizer
# opt.vgg_unnorm = VggUnNormalizer().to('cuda')
from DGPT.Utils.Preprocess import RGB2YUV, YUV2RGB
opt.rgb2yuv = RGB2YUV('cuda')
opt.yuv2rgb = YUV2RGB('cuda')
# prepare mask
face_mask = torch.ones(1, 1, 512, 512).to('cuda')
face_mask *= 0.1