def get_gt_mask_temp(image_filename, args):
mask_final_path = os.path.join(args.data_path, 'occ_mask_final', image_filename)
gt_path_abs = get_gt_path(image_filename, args.dataset)
gt_depth = cv_io.read(gt_path_abs)[192:192+128,192:192+160]
final_soft = cv_io.read(mask_final_path).astype(np.float32) / 255.
if len(final_soft.shape) > 2: final_soft = final_soft[...,0]
final = np.zeros_like(final_soft)
final[final_soft>0.5] = 1.0
final[gt_depth<0.1] = 0.5
gt_mask_path = os.path.join('tmp', 'gt_mask_'+image_filename).replace('.dpt', '.png')
cv_io.save(gt_mask_path, (final*255).astype(np.uint8))
return gt_mask_path
def get_dif_map(est_path, gt_path, tol=0.2):
est = cv_io.read(est_path)
gt = cv_io.read(gt_path)
m = np.min(gt)
M = np.max(gt)
gt_nromal = (gt - m) / (M - m) * 255
R = gt_nromal.copy()
G = gt_nromal.copy()
B = gt_nromal.copy()
R [np.abs(est-gt)>tol] = 255
G [np.abs(est-gt)>tol] = 0
B [np.abs(est-gt)>tol] = 0
R [gt < 1e-3] = 0
G [gt < 1e-3] = 0
B [gt < 1e-3] = 0
error = np.stack([R,G,B],-1)
cv_io.save('tmp/x.png', error.astype(np.uint8))
return 'tmp/x.png'
def score(image_filename, metric, args):
mask_final_path = os.path.join(args.data_path, 'occ_mask_final', image_filename)
mask_gt_path = os.path.join(args.data_path, 'occ_mask_gt', image_filename)
gt_path = get_gt_path(image_filename, args.dataset)
gt_depth = cv_io.read(gt_path)[192:192+128,192:192+160]
final_soft = cv_io.read(mask_final_path).astype(np.float32) / 255.
if len(final_soft.shape) > 2: final_soft = final_soft[...,0]
final = np.zeros_like(final_soft)
final[final_soft>0.5] = 1.0
gt = cv_io.read(mask_gt_path).astype(np.float32) / 255.
if len(gt.shape) > 2: gt = gt[...,0]
final = final[gt_depth>0.1]
gt = gt[gt_depth>0.1]
final_DR = np.sum(final!=gt) / np.prod(gt.shape)
return final_DR
def ROI_box(rgb_file, args, thickness=3):
rgb = cv_io.read(rgb_file)
th = thickness
x = 192; y = 192
rgb[x-th:x+th,y-th:y+160+th,:] = [0,255,0]
rgb[x+128-th:x+128+th,y-th:y+160+th,:] = [0,255,0]
rgb[x-th:x+128+th,y-th:y+th,:] = [0,255,0]
rgb[x-th:x+128+th,y+160-th:y+160+th,:] = [0,255,0]
temp_save = os.path.join ('tmp', 'rgb_'+rgb_file.split(os.sep)[-1])
cv_io.save(temp_save, rgb)
return temp_save
def get_gt_temp(image, args, inverse=True):
gt_path_abs = get_gt_path(image, args.dataset)
gt = cv_io.read(gt_path_abs)
gt_norm = 1 / gt if inverse else gt.copy()
gt_norm[gt<1e-3] = 0
m = np.min(gt_norm)
M = np.max(gt_norm)
gt_norm = (gt_norm - m) / (M - m) * 255
# gt_norm[gt<1e-3] = 255 / 2
gt_path = os.path.join('tmp', 'gt_'+image).replace('.dpt', '.png')
cv_io.save(gt_path, gt_norm.astype(np.uint8))
return gt_path
def ROI_box(gt, image, h, w, M, thickness=3):
rgb = cv_io.read(image)
rgb = draw_box(rgb, h, w)
mask = np.zeros_like(gt)
mask[gt < M] = 0
mask[gt > M] = 255
mask[gt < 0.01] = 255 // 2
mask_color = mask[:, :, np.newaxis]
mask_color = np.concatenate(
(mask_color, np.zeros_like(mask_color), np.zeros_like(mask_color)),
axis=2)
mask_color = 0.8 * rgb + 0.2 * mask_color
temp_save = os.path.join('examples/tmp',
'mask_ROI_' + image.replace(os.sep, '_')).replace(
'.dpt', '.png')
cv_io.save(temp_save, mask_color.astype(np.uint8))
return temp_save
def get_planes(gt, image, args, examples=4):
gap_points = []
th = 250
while len(gap_points) < 8:
gap_points = plane_finder.get_points(gt * 1000,
None,
args.n,
args.k,
th=th,
f=args.f)
th = th / 2
line_normals, Ms, selected_points = plane_finder.find_lines(
gap_points,
number_of_lines=args.number_of_planes,
method='ransac',
plane_th=50,
s=2)
img = cv_io.read(image)
mask_paths = list()
masks = list()
for i in range(examples):
line_normal = line_normals[i]
M = Ms[i]
N = np.asarray([0, line_normal[2], -line_normal[1]])
if np.linalg.norm(N) == 0:
N = np.asarray([0, 0, -1])
else:
N = N / np.linalg.norm(N)
N = np.reshape(N, [1, 1, 3])
M = np.reshape(M, [1, 1, 3])
offset = np.sum(-M * N)
N = N / np.linalg.norm(N) * np.sign(-offset)
offset *= np.sign(-offset)
x = np.linspace(0, 1, gt.shape[0])
y = np.linspace(0, 1, gt.shape[1])
xv, yv = np.meshgrid(x, y, indexing='ij')
xv = (xv[:, :, np.newaxis] - 0.5) * gt.shape[0] / args.f
yv = (yv[:, :, np.newaxis] - 0.5) * gt.shape[1] / args.f
X = np.concatenate((xv, yv, np.ones_like(xv)), axis=2)
Z = np.sum(N * X, axis=2) * gt * 1000 + offset
mask = np.zeros_like(gt)
mask[Z > 0] = 0
mask[Z < 0] = 255
mask[gt < 0.01] = 255 // 2
corner_flag = True
j = 0
best_score = 1.0
while corner_flag:
j += 1
h_tmp = np.random.randint(Z.shape[0] - 128)
w_tmp = np.random.randint(Z.shape[1] - 160)
ROI = Z[h_tmp:h_tmp + 128, w_tmp:w_tmp + 160]
score = np.abs(np.mean(ROI) / 255 - 0.5)
if score < best_score:
best_score = score
h = h_tmp
w = w_tmp
if score < 0.25 or j > 100:
corner_flag = False
h = h_tmp
w = w_tmp
mask_color = 255 - mask[:, :, np.newaxis]
mask_color = np.concatenate(
(np.zeros_like(mask_color), mask_color, np.zeros_like(mask_color)),
axis=2)
mask_color = 0.8 * img + 0.2 * mask_color
mask_color = draw_box(mask_color, h, w, color=[0, 128, 0])
mask_path = os.path.join(
'examples/tmp',
'mask_' + str(i) + '_' + image.replace(os.sep, '_')).replace(
'.dpt', '.png')
cv_io.save(mask_path, mask_color.astype(np.uint8))
mask_paths.append(mask_path)
masks.append(mask)
return mask_paths, masks
items = ['Examples are sorted based on the DR metric']
items.append('Green rectangular is Region of Interest (ROI)')
items.append('Final mask is a soft mask')
items.append("Slide's title is combination of the image name and its DR values")
items.append('Python generator code is available on \\href{https://github.com/saeid-h/Latex-Slides}{Github}')
intro_slide = Items(items)
slide = Frame('Introduction', intro_slide)
presentation.add_frames(slide)
for i, image_tuple in enumerate(images):
image, score = image_tuple
title = image.split(os.sep)[-1].split('.')[0].replace('_', '\_') + ' ::: ' + str(round(score*100,2)) + '\%'
scale = 0.15 if args.dataset == 'nyu' else 0.18
gt_path_abs = get_gt_path(image, args.dataset)
gt = cv_io.read(gt_path_abs)
rgb_path = os.path.join(args.data_path, 'rgb', image)
rgb_path = ROI_box(rgb_path, args).replace('_', '\string_')
rgb_img = Graphics(scale, path=rgb_path)
gt_path = get_gt_temp(image, args)
gt_path = ROI_box_gt(gt, rgb_path).replace('_', '\string_')
gt_img = Graphics(scale, path=gt_path)
mask_gt_path = get_gt_mask_temp(image, args)
mask_gt_img = Graphics(scale=0.58, path=mask_gt_path)
table_data = [[rgb_img, gt_img, mask_gt_img]]
table_data += [['RGB', 'GT', 'GT Mask']]
