def vis_mask(sub_dir, split, num=10, save_dir='', num_parts=14):
im_files = get_im_names(osp.join(root_dir, 'images', split),
pattern='*.jpg',
return_np=False,
return_path=True)
mask_files = get_im_names(osp.join(root_dir, sub_dir, split),
pattern='*.png',
return_np=False,
return_path=True)
im_files.sort()
mask_files.sort()
assert len(im_files) == len(mask_files)
assert all([
osp.basename(im_f)[:-4] == osp.basename(mask_f)[:-4]
for im_f, mask_f in zip(im_files, mask_files)
])
for im_f, mask_f in list(zip(im_files, mask_files))[:num]:
for vis_func, save_sub_dir in zip(
[vis_mask_, vis_mask_each_part_],
['all_on_one_im', 'one_im_per_mask']):
vis_func(
read_im(im_f,
convert_rgb=True,
resize_h_w=None,
transpose=False),
read_im(mask_f,
convert_rgb=False,
resize_h_w=None,
transpose=False),
osp.join(save_dir, save_sub_dir, osp.basename(im_f)),
num_parts)
def vis_parts(im_dir, ann):
# All keys:
# 'category_id', 'id', 'image_id',
# 'iscrowd', 'area',
# 'num_keypoints', 'keypoints',
# 'bbox', 'segmentation', 'dp_masks',
# 'dp_x', 'dp_y', 'dp_I', 'dp_U', 'dp_V'
# print(ann.keys())
part_mask = GetDensePoseMask(ann['dp_masks'])
bbr = np.array(ann['bbox']).astype(int) # the box.
################
im = read_im(osp.join(im_dir, im_id_to_name[ann['image_id']]),
resize_h_w=None,
transpose=False)
x1, y1, x2, y2 = bbr[0], bbr[1], bbr[0] + bbr[2], bbr[1] + bbr[3]
x2 = min([x2, im.shape[1]])
y2 = min([y2, im.shape[0]])
################
MaskIm = cv2.resize(part_mask, (int(x2 - x1), int(y2 - y1)),
interpolation=cv2.INTER_NEAREST)
MaskBool = np.tile((MaskIm == 0)[:, :, np.newaxis], [1, 1, 3])
# Replace the visualized mask image with I_vis.
Mask_vis = cv2.applyColorMap((MaskIm * 15).astype(np.uint8),
cv2.COLORMAP_PARULA)[:, :, :]
Mask_vis[MaskBool] = im[y1:y2, x1:x2, :][MaskBool]
im = im.copy().astype(np.float)
im[y1:y2, x1:x2, :] = im[y1:y2, x1:x2, :] * 0.3 + Mask_vis * 0.7
return im
def fuse_parts_(mask_f, new_mask_f):
ori_mask = read_im(mask_f,
convert_rgb=False,
resize_h_w=None,
transpose=False)
mask = ori_mask.copy()
for m in mapping:
mask[ori_mask == m[0]] = m[1]
save_im(mask, new_mask_f, transpose=False, check_bound=False)
def check_mask_value_(mask_file):
mask = read_im(mask_file,
convert_rgb=False,
resize_h_w=None,
transpose=False)
assert len(mask.shape) == 2, "{} mask.shape is {}".format(
mask_file, mask.shape)
assert mask.min() >= 0, "{} mask.min is {}".format(mask_file, mask.min())
assert mask.max() <= 14, "{} mask.max is {}".format(mask_file, mask.max())
def check_mask_value(split):
"""To make sure saving and loading mask images are correct."""
mask_files = get_im_names(osp.join(save_root_dir, mask_dir, split),
pattern='*.png',
return_np=False,
return_path=True)
for mf in mask_files:
mask = read_im(mf, convert_rgb=False, resize_h_w=None, transpose=False)
check_mask_value_(mask)
def vis_mask(split, num=10, save_dir=''):
im_files = get_im_names(osp.join(save_root_dir, 'images', split),
pattern='*.jpg',
return_np=False,
return_path=True)
mask_files = get_im_names(osp.join(save_root_dir, mask_dir, split),
pattern='*.png',
return_np=False,
return_path=True)
im_files.sort()
mask_files.sort()
assert len(im_files) == len(mask_files)
assert all([
osp.basename(im_f)[:-4] == osp.basename(mask_f)[:-4]
for im_f, mask_f in zip(im_files, mask_files)
])
for im_f, mask_f in list(zip(im_files, mask_files))[:num]:
vis_mask_(
read_im(im_f, convert_rgb=True, resize_h_w=None, transpose=False),
read_im(mask_f,
convert_rgb=False,
resize_h_w=None,
transpose=False), osp.join(save_dir, osp.basename(im_f)))
def vis_fg(im_dir, ann):
# All keys:
# 'category_id', 'id', 'image_id',
# 'iscrowd', 'area',
# 'num_keypoints', 'keypoints',
# 'bbox', 'segmentation', 'dp_masks',
# 'dp_x', 'dp_y', 'dp_I', 'dp_U', 'dp_V'
# print(ann.keys())
bbox = np.array(ann['bbox']).astype(int) # the box.
################
im = read_im(osp.join(im_dir, im_id_to_name[ann['image_id']]),
resize_h_w=None,
transpose=False)
x1, y1, x2, y2 = bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]
x2 = min([x2, im.shape[1]])
y2 = min([y2, im.shape[0]])
################
# Example:
# => im.shape (428, 640, 3)
# => MaskIm.shape (428, 640)
# => MaskIm.dtype uint8
# => np.unique(MaskIm) [0 1]
# # MaskIm.shape (428, 640)
# # MaskIm.dtype uint8
# # np.unique(MaskIm) [0 1]
# The same resolution as whole image.
MaskIm = annToMask(ann, im.shape[0], im.shape[1])
print('=> im.shape', im.shape)
print('=> MaskIm.shape', MaskIm.shape)
print('=> MaskIm.dtype', MaskIm.dtype)
print('=> np.unique(MaskIm)', np.unique(MaskIm))
if MaskIm.shape != im.shape[:2]:
MaskIm = cv2.resize(MaskIm, (im.shape[1], im.shape[0]),
interpolation=cv2.INTER_NEAREST)
print('# MaskIm.shape', MaskIm.shape)
print('# MaskIm.dtype', MaskIm.dtype)
print('# np.unique(MaskIm)', np.unique(MaskIm))
################
MaskBool = np.tile((MaskIm == 0)[:, :, np.newaxis], [1, 1, 3])
# Replace the visualized mask image with I_vis.
Mask_vis = MaskIm[..., np.newaxis] * np.array([[[255, 0, 0]]])
Mask_vis[MaskBool] = im[MaskBool]
im = im.copy().astype(np.float)
im = im * 0.3 + Mask_vis * 0.7
return im
def transform(ann, im_dir, save_im_dir, save_mask_dir, split_file):
"""input ann: loaded json"""
may_make_dir(osp.dirname(split_file))
fid = open(split_file, 'w')
im_id_to_name = get_im_id_to_name(ann)
anns = ann['annotations']
i = 0
for ann in anns:
if not 'dp_masks' in ann:
continue
bbox = np.array(ann['bbox']).astype(int)
im = read_im(osp.join(im_dir, im_id_to_name[ann['image_id']]),
resize_h_w=None,
transpose=False)
x1, y1, x2, y2 = bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]
x1 = max([x1, 0])
y1 = max([y1, 0])
x2 = min([x2, im.shape[1]])
y2 = min([y2, im.shape[0]])
mask = GetDensePoseMask(ann['dp_masks']).astype(np.uint8)
mask = cv2.resize(mask, (x2 - x1, y2 - y1),
interpolation=cv2.INTER_NEAREST)
assert mask.min() >= 0, "mask.min is {}".format(mask.min())
assert mask.max() <= 14, "mask.max is {}".format(mask.max())
im = im[y1:y2, x1:x2, :]
save_im_name = im_id_to_name[ann['image_id']][:-4] + '_' + str(
ann['id']) + '.jpg'
save_im_path = osp.join(save_im_dir, save_im_name)
save_im(im, save_im_path, transpose=False, check_bound=False)
save_mask_name = im_id_to_name[ann['image_id']][:-4] + '_' + str(
ann['id']) + '.png'
save_mask_path = osp.join(save_mask_dir, save_mask_name)
save_im(mask, save_mask_path, transpose=False, check_bound=False)
rel_im_path = '/'.join(save_im_path.split('/')[-3:])
rel_mask_path = '/'.join(save_mask_path.split('/')[-3:])
fid.write('{}\t{}\n'.format(rel_im_path, rel_mask_path))
i += 1
fid.close()
return len(anns), i
def transform(ann,
im_dir,
save_im_dir,
save_mask_dir,
split_file,
to_save_im=False):
"""
Args:
ann: loaded Densepose annotation json
to_save_im: If your cropped images have been saved when preparing part parsing, this can be False
"""
may_make_dir(osp.dirname(split_file))
fid = open(split_file, 'w')
im_id_to_name = get_im_id_to_name(ann)
anns = ann['annotations']
i = 0
n_dp = 0
n_fg = 0
for ann in anns:
i += 1
if not 'dp_masks' in ann:
continue
n_dp += 1
if not 'segmentation' in ann:
continue
n_fg += 1
bbox = np.array(ann['bbox']).astype(int)
im = read_im(osp.join(im_dir, im_id_to_name[ann['image_id']]),
resize_h_w=None,
transpose=False)
x1, y1, x2, y2 = bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]
x1 = max([x1, 0])
y1 = max([y1, 0])
x2 = min([x2, im.shape[1]])
y2 = min([y2, im.shape[0]])
# The same resolution as whole image.
MaskIm = annToMask(ann, im.shape[0], im.shape[1])
assert len(MaskIm.shape) == 2, "len(MaskIm.shape) {}".format(
len(MaskIm.shape))
assert MaskIm.shape == im.shape[:
2], "MaskIm.shape {}, im.shape {}".format(
MaskIm.shape, im.shape)
# print('=> im.shape', im.shape)
# print('=> MaskIm.shape', MaskIm.shape)
# print('=> MaskIm.dtype', MaskIm.dtype)
# print('=> np.unique(MaskIm)', np.unique(MaskIm))
# The same resolution as bbox
MaskIm = MaskIm[y1:y2, x1:x2]
# print('=> MaskIm.shape', MaskIm.shape)
check_mask_value_(MaskIm)
save_im_name = im_id_to_name[ann['image_id']][:-4] + '_' + str(
ann['id']) + '.jpg'
save_im_path = osp.join(save_im_dir, save_im_name)
if to_save_im:
im = im[y1:y2, x1:x2, :]
save_im(im, save_im_path, transpose=False, check_bound=False)
save_mask_name = im_id_to_name[ann['image_id']][:-4] + '_' + str(
ann['id']) + '_fg' + '.png'
save_mask_path = osp.join(save_mask_dir, save_mask_name)
save_im(MaskIm, save_mask_path, transpose=False, check_bound=False)
rel_im_path = '/'.join(save_im_path.split('/')[-3:])
rel_mask_path = '/'.join(save_mask_path.split('/')[-3:])
fid.write('{}\t{}\n'.format(rel_im_path, rel_mask_path))
if i % 200 == 0:
print('{}/{} Done'.format(i, len(anns)))
fid.close()
return len(anns), n_dp, n_fg
# Including Background class
nclass = 8
# infer_im_path_mask_path_file = 'datasets/cuhk03/detected/infer_im_path_mask_path.txt'
infer_im_path_mask_path_file = 'datasets/duke/infer_im_path_mask_path.txt'
lines = read_lines(infer_im_path_mask_path_file, strip=True)
im_paths, mask_paths = zip(*[l.split('\t') for l in lines])
rand_inds = np.random.permutation(range(len(im_paths)))
im_paths = np.array(im_paths)[rand_inds]
mask_paths = np.array(mask_paths)[rand_inds]
num_vis_ims = 64
im_grid_list = []
for i in range(num_vis_ims):
im = read_im(im_paths[i],
convert_rgb=True,
resize_h_w=None,
transpose=False)
mask = read_im(mask_paths[i],
convert_rgb=False,
resize_h_w=None,
transpose=False)
vis_one_im(im,
mask,
nclass,
resize_h_w=(128, 64),
im_grid_list=im_grid_list)
n_cols = 8
n_rows = int(np.ceil(len(im_grid_list) / n_cols))
vis_im = make_im_grid(im_grid_list, n_rows, n_cols, 8, 255)
save_im(vis_im,
'exp/vis_coco_part_pred/cuhk03_detected/vis_im.jpg',
