def gao(idx):
# if os.path.exists(test_label_dir.joinpath('%s.torch' % idx)):
# return
anno_path = anno_dir.joinpath('%s_%s_keypoints.json' % (args.name, idx))
anno = json.load(open(str(anno_path)))['people']
if len(anno) == 0:
print("warnings: %s no people" % idx)
return [-1, -1, -1, -1]
anno = anno[0]
for key in anno:
anno[key] = np.array(anno[key]).reshape(-1, 3).astype(np.float)
anno[key][:, 0] -= 420
anno[key][:, :2] = (anno[key][:, :2] / 1080. * 512).clip(min=0)
x = anno['pose_keypoints_2d'][:, :2]
s = np.linalg.norm(x[1, :] - x[8, :])
y = max(x[21, 1], x[24, 1])
if y > 10 and x[1, :].min() > 5 and x[8, :].min() > 5:
w = x[:, 0].max() - x[:, 0].min()
b = (y - ymin) / (ymax - ymin) * (fmax - fmin) + fmin
l = (y - ymin) / (ymax - ymin) * (tmax / smax -
tmin / smin) + tmin / smin
left = w * (1 - l) * x[:, 1].min() / (512 - w)
for key in anno:
anno[key] *= l
d = max(x[21, 1], x[24, 1]) - b
# print(l, d)
for key in anno:
anno[key][:, 0] += left
anno[key][:, 1] -= d
# print(max(x[21,1], x[24,1]) - b)
# img_path = img_dir.joinpath('%s.png'%idx)
# img = cv2.imread(str(img_path))[:, 420: -420]
# img = cv2.resize(img, (512, 512))
# cv2.imwrite(str(test_image_dir.joinpath('%s.png'%idx)), img)
# label = create_label_full((512, 512), anno)
# s = label.max(axis = 2)[:,:, np.newaxis]
# fig = plt.figure(1)
# ax = fig.add_subplot(111)
# ax.imshow((img * .8 + s * 255 * .2 ).astype(np.uint8))
# ax.imshow((s[:,:, 0] * 255).astype(np.uint8))
# plt.show()
# label = torch.tensor(label).byte()
# label_path = test_label_dir.joinpath('%s.torch'% idx)
# torch.save(label, str(label_path))
# print(str(test_image_dir.joinpath('%s.png'%idx)))
# ================ Crop Face=====================
face = anno['face_keypoints_2d']
if face[:, 2].min() < 0.001:
print(face[:, 2].min())
return [-1, -1, -1, -1]
minx, maxx = int(max(face[:, 1].min() - 20,
0)), int(min(face[:, 1].max() + 10, 512))
miny, maxy = int(max(face[:, 0].min() - 15,
0)), int(min(face[:, 0].max() + 15, 512))
face[:, 0] = (face[:, 0] - miny) / (maxy - miny + 1.) * 128.
face[:, 1] = (face[:, 1] - minx) / (maxx - minx + 1.) * 128.
face_label = create_face_label((128, 128), face)
# fig = plt.figure(1)
# ax = fig.add_subplot(111)
# s = face_label.max(axis = 2, keepdims = False)
# print(s.shape)
# ax.imshow(s)
# plt.show()
face_label = torch.tensor(face_label).byte()
face_label_dir = test_face_label_dir.joinpath('%s.torch' % idx)
torch.save(face_label, str(face_label_dir))
return [minx, maxx, miny, maxy]
def gao(idx):
# if os.path.exists(train_head_dir.joinpath('%s.png' % idx)):
# return
print(idx)
anno_path = anno_dir.joinpath('%s_%s_keypoints.json' % (args.name, idx))
anno = json.load(open(str(anno_path)))['people']
if len(anno) == 0:
print("warnings: %s no people" % idx)
return [-1, -1, -1, -1]
anno = anno[0]
anno.pop("person_id")
# print(np.array(anno['pose_keypoints_2d']).reshape(-1, 3))
for key in anno:
anno[key] = np.array(anno[key]).reshape(-1, 3)
anno[key][:, 0] -= 420
# print(anno[key][:, 0].min())
anno[key][:, :2] = (anno[key][:, :2] / 1080. * 512).clip(min=0)
anno['person_id'] = temp_dict['person_id']
# print(anno['pose_keypoints_2d'][:, :2][:, ::-1])
img_path = img_dir.joinpath('%s.png' % idx)
img = cv2.imread(str(img_path))[:, 420:-420]
img = cv2.resize(img, (512, 512))
cv2.imwrite(str(train_img_dir.joinpath(idx + '.png')), img)
label = create_label_full((512, 512), anno)
# s = label.max(axis = 2)[:,:, np.newaxis]
# fig = plt.figure(1)
# ax = fig.add_subplot(111)
# ax.imshow((img * .8 + s * 255 * .2 ).astype(np.uint8))
# ax.imshow((s[:,:, 0] * 255).astype(np.uint8))
# plt.show()
label = torch.tensor(label).byte()
label_path = train_label_dir.joinpath('%s.torch' % idx)
torch.save(label, str(label_path))
# ===================== Crop Face ==============================
face = anno['face_keypoints_2d']
if face[:, 2].min() < 0.001:
print(face[:, 2].min())
return [-1, -1, -1, -1]
minx, maxx = int(max(face[:, 1].min() - 20,
0)), int(min(face[:, 1].max() + 10, 512))
miny, maxy = int(max(face[:, 0].min() - 15,
0)), int(min(face[:, 0].max() + 15, 512))
# print(minx, maxx , miny, maxy )
face_img = img[minx:maxx + 1, miny:maxy + 1, :]
face_img = cv2.resize(face_img, (128, 128))
cv2.imwrite(str(train_face_image_dir.joinpath('%s.png' % idx)), face_img)
face[:, 0] = (face[:, 0] - miny) / (maxy - miny + 1.) * 128.
face[:, 1] = (face[:, 1] - minx) / (maxx - minx + 1.) * 128.
face_label = create_face_label((128, 128), face)
# s = face_label.max(axis = 2)[:,:, np.newaxis]
# fig = plt.figure(1)
# ax = fig.add_subplot(111)
# ax.imshow((s[:,:, 0] * 255).astype(np.uint8))
# plt.show()
face_label = torch.tensor(face_label).byte()
face_label_dir = train_face_label_dir.joinpath('%s.torch' % idx)
torch.save(face_label, str(face_label_dir))
return [minx, maxx, miny, maxy]