def __getitem__(self, item):
img_target_path_1 = self.img_target_paths_1[item]
img_target_path_2 = self.img_target_paths_2[item]
img_target_1 = Image.open(img_target_path_1).convert('RGB')
img_target_2 = Image.open(img_target_path_2).convert('RGB')
img_target_1 = img_target_1.resize(
[self.opt.loadSize_v[0], self.opt.loadSize_v[1]], Image.BICUBIC)
img_target_2 = img_target_2.resize(
[self.opt.loadSize_v[0], self.opt.loadSize_v[1]], Image.BICUBIC)
lab_target_path = self.lab_target_paths[item]
lab_target = read_flo_file(lab_target_path)
lab_target_h, lab_target_w = lab_target.shape[0], lab_target.shape[1]
lab_target = resample(lab_target,
[self.opt.loadSize_v[1], self.opt.loadSize_v[0]])
lab_target = torch.from_numpy(lab_target.transpose(2, 0, 1))
lab_target[0] = lab_target[0] * self.opt.loadSize_v[0] / lab_target_w
lab_target[1] = lab_target[1] * self.opt.loadSize_v[1] / lab_target_h
img_target_1 = self.transform_no_augment(img_target_1)
img_target_2 = self.transform_no_augment(img_target_2)
# print('img_source:',img_source_1.size())
return {
'img_target': [img_target_1, img_target_2],
'lab_target': lab_target,
'img_target_paths': [img_target_path_1, img_target_path_2],
'lab_target_path': lab_target_path
}
def load_opticalflows_flo(flow_dir, im_names, flo_im_ext):
forwardflows = []
backwardflows = []
fw_flow_arr = []
bw_flow_arr = []
fw_flow_dir = os.path.join(flow_dir, 'fw')
bw_flow_dir = os.path.join(flow_dir, 'bw')
for flow_id in xrange(len(im_names) - 1):
flow_name = im_names[flow_id] + flo_im_ext
fw_flow_path = os.path.join(fw_flow_dir, flow_name)
bw_flow_path = os.path.join(bw_flow_dir, flow_name)
fw_flow = read_flo_file(fw_flow_path) ##fw_flow.shape=(1080, 1920, 2)
bw_flow = read_flo_file(bw_flow_path)
forwardflows.append(fw_flow)
backwardflows.append(bw_flow)
fw_flow_arr = np.asarray(forwardflows) ## list to array
bw_flow_arr = np.asarray(backwardflows)
# if flow_id==0:
# visualize_flow(fw_flow)
print 'load', len(forwardflows), 'forward and backward flow files done...'
return fw_flow_arr, bw_flow_arr
def __getitem__(self, item):
index = random.randint(0, self.img_target_size - 1)
img_source_path_1 = self.img_source_paths_1[item %
self.img_source_size]
img_source_path_2 = self.img_source_paths_2[item %
self.img_source_size]
if self.opt.dataset_mode == 'paired':
img_target_path_1 = self.img_target_paths_1[item %
self.img_target_size]
img_target_path_2 = self.img_target_paths_2[item %
self.img_target_size]
elif self.opt.dataset_mode == 'unpaired':
img_target_path_1 = self.img_target_paths_1[index]
img_target_path_2 = self.img_target_paths_2[index]
else:
raise ValueError('Data mode [%s] is not recognized' %
self.opt.dataset_mode)
img_source_1 = Image.open(img_source_path_1).convert('RGB')
img_source_2 = Image.open(img_source_path_2).convert('RGB')
img_target_1 = Image.open(img_target_path_1).convert('RGB')
img_target_2 = Image.open(img_target_path_2).convert('RGB')
img_source_1 = img_source_1.resize(
[self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)
img_source_2 = img_source_2.resize(
[self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)
img_target_1 = img_target_1.resize(
[self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)
img_target_2 = img_target_2.resize(
[self.opt.loadSize[0], self.opt.loadSize[1]], Image.BICUBIC)
if self.opt.isTrain:
lab_source_path = self.lab_source_paths[item %
self.lab_source_size]
# print(lab_source_path)
lab_source = read_flo_file(lab_source_path)
# flowtool = flow2img.Flow()
# flow = flowtool._readFlow(lab_source_path)
# print('flow:', flow.shape)
# img = flowtool._flowToColor(lab_source.transpose(2,0,1))
# img = Image.fromarray(img).save('/home/lyc/Desktop/flow.png')
lab_source_h, lab_source_w = lab_source.shape[0], lab_source.shape[
1]
lab_source = resample(lab_source,
[self.opt.loadSize[1], self.opt.loadSize[0]])
# print(lab_source.shape)
lab_source = torch.from_numpy(lab_source.transpose(2, 0, 1))
lab_source[0] = lab_source[0] * self.opt.loadSize[0] / lab_source_w
lab_source[1] = lab_source[1] * self.opt.loadSize[1] / lab_source_h
lab_target_path = self.lab_target_paths[item %
self.lab_target_size]
lab_target = read_flo_file(lab_target_path)
lab_target_h, lab_target_w = lab_target.shape[0], lab_target.shape[
1]
lab_target = resample(lab_target,
[self.opt.loadSize[1], self.opt.loadSize[0]])
lab_target = torch.from_numpy(lab_target.transpose(2, 0, 1))
lab_target[0] = lab_target[0] * self.opt.loadSize[0] / lab_target_w
lab_target[1] = lab_target[1] * self.opt.loadSize[1] / lab_target_h
# print('lab_source:',lab_source.size())
img_source_1 = self.transform_augment(img_source_1)
img_source_2 = self.transform_augment(img_source_2)
img_target_1 = self.transform_no_augment(img_target_1)
img_target_2 = self.transform_no_augment(img_target_2)
# print('img_source:',img_source_1.size())
return {
'img_source': [img_source_1, img_source_2],
'img_target': [img_target_1, img_target_2],
'lab_source': lab_source,
'lab_target': lab_target,
'img_source_paths': [img_source_path_1, img_source_path_2],
'img_target_paths': [img_target_path_1, img_target_path_2],
'lab_source_path': lab_source_path,
'lab_target_path': lab_target_path
}
else:
img_source = self.transform_augment(img_source)
img_target = self.transform_no_augment(img_target)
return {
'img_source': img_source,
'img_target': img_target,
'img_source_paths': img_source_path,
'img_target_paths': img_target_path,
}
img_seq[seq]) + "/frame_" + "0" + str(imgNum).zfill(3) + ".png"
img2Path = imgPathPrefix + str(
img_seq[seq]) + "/frame_" + "0" + str(imgNum +
1).zfill(3) + ".png"
flowPath = flowPathPrefix + "/flow" + str(
img_seq[seq]) + "/frame_" + "0" + str(imgNum).zfill(3) + ".flo"
flowPath_invalid = flowPathPrefix + "/invalid" + str(
img_seq[seq]) + "/frame_" + "0" + str(imgNum).zfill(3) + ".png"
flowPath_occ = flowPathPrefix + "/occlusions" + str(
img_seq[seq]) + "/frame_" + "0" + str(imgNum).zfill(3) + ".png"
img1 = flowlib.read_image(img1Path)
img2 = flowlib.read_image(img2Path)
height = img1.shape[0]
width = img1.shape[1]
flow = flowlib.read_flo_file(flowPath)
u = flow[:, :, 0]
v = flow[:, :, 1]
invalid, _ = flowlib.read_image(flowPath_invalid)
occ, _ = flowlib.read_image(flowPath_occ)
for i in range(halfPatchSize, height - halfPatchSize, 15):
for j in range(halfPatchSize, width - halfPatchSize, 15):
if invalid[i, j] == 0 and occ[i, j] == 0:
if (i + int(v[i, j] - 1) - halfPatchSize < 0
or i + int(v[i, j] + 1) + halfPatchSize >=
height):
continue