def hookTrainData(self, sampleIdxs):
assert len(sampleIdxs) > 0, 'we need a non-empty batch list'
input_list, flow_list = [], []
for idx in sampleIdxs:
img_list = self.trainList[idx]
multi_input = []
multi_flow = []
for time_idx in xrange(self.time_step):
imgData = cv2.imread(
os.path.join(self.img_path, img_list[time_idx]),
cv2.IMREAD_COLOR)
multi_input.append(
np.expand_dims(
cv2.resize(imgData,
(self.image_size[1], self.image_size[0])),
0))
# We have self.time_step images, but self.time_step - 1 flows.
if time_idx != self.time_step - 1:
flow = utils.readFlow(
os.path.join(self.data_path, 'training', "flow",
(img_list[time_idx][:-4] + ".flo")))
multi_flow.append(np.expand_dims(flow, 0))
input_list.append(np.concatenate(multi_input, axis=3))
flow_list.append(np.concatenate(multi_flow, axis=3))
return np.concatenate(input_list, axis=0), np.concatenate(flow_list,
axis=0)
def hookTrainData(self, sampleIdxs):
assert len(sampleIdxs) > 0, 'we need a non-empty batch list'
source_list, target_list, flow_gt = [], [], []
for idx in sampleIdxs:
frameID = self.trainList[idx]
prev_img = frameID + "_img1.ppm"
next_img = frameID + "_img2.ppm"
source = cv2.imread(os.path.join(self.img_path, prev_img),
cv2.IMREAD_COLOR)
target = cv2.imread(os.path.join(self.img_path, next_img),
cv2.IMREAD_COLOR)
# print source.shape
flow = utils.readFlow(
os.path.join(self.img_path, (frameID + "_flow.flo")))
# print flow.shape
source = cv2.resize(source,
(self.image_size[1], self.image_size[0]))
source_list.append(np.expand_dims(source, 0))
target = cv2.resize(target,
(self.image_size[1], self.image_size[0]))
target_list.append(np.expand_dims(target, 0))
flow_gt.append(np.expand_dims(flow, 0))
return np.concatenate(source_list, axis=0), np.concatenate(
target_list, axis=0), np.concatenate(flow_gt, axis=0)
def hookTrainData(self, sampleIdxs):
assert len(sampleIdxs) > 0, 'we need a non-empty batch list'
source_list, target_list, flow_gt = [], [], []
for idx in sampleIdxs:
img_pair = self.trainList[idx]
prev_img = img_pair[0]
next_img = img_pair[1]
source = cv2.imread(os.path.join(self.img_path, prev_img),
cv2.IMREAD_COLOR)
target = cv2.imread(os.path.join(self.img_path, next_img),
cv2.IMREAD_COLOR)
flow = utils.readFlow(
os.path.join(self.data_path, 'training', "flow",
(prev_img[:-4] + ".flo")))
if self.is_crop:
source = cv2.resize(source,
(self.crop_size[1], self.crop_size[0]))
target = cv2.resize(target,
(self.crop_size[1], self.crop_size[0]))
source_list.append(np.expand_dims(source, 0))
target_list.append(np.expand_dims(target, 0))
flow_gt.append(np.expand_dims(flow, 0))
return np.concatenate(source_list, axis=0), np.concatenate(
target_list, axis=0), np.concatenate(flow_gt, axis=0)
# sess.run(init)
saver.restore(sess, model_path)
test_dir = './middlebury_data/Grove2'
# img1 = read_img(test_dir + '/' + 'frame10.png')[0:mdl_height, 0:mdl_width, :]
# img2 = read_img(test_dir + '/' + 'frame11.png')[0:mdl_height, 0:mdl_width, :]
# edge = np.expand_dims(read_img(test_dir + '/' + 'frame10_edge.png')[0:mdl_height, 0:mdl_width], 2)
# miss = np.expand_dims(np.mean(np.abs(img1-img2), axis=2), 2)
# flow_gt = readFlow(test_dir + '/' + 'flow10.flo')[0:mdl_height, 0:mdl_width, :]
img1 = read_img(test_dir + '/' + 'frame10.png')
img2 = read_img(test_dir + '/' + 'frame11.png')
edge = np.expand_dims(read_img(test_dir + '/' + 'frame10_edge.png'), 2)
miss = np.expand_dims(np.mean(np.abs(img1 - img2), axis=2), 2)
flow_gt = readFlow(test_dir + '/' + 'flow10.flo')
H = img1.shape[0]
W = img1.shape[1]
flow = np.zeros(shape=(H, W, 2), dtype=np.float32)
for i in range(int(floor(H / float(mdl_height)))):
for j in range(int(floor(W / float(mdl_width)))):
start_y = i * mdl_height
start_x = j * mdl_width
img1_tile = img1[start_y:start_y + mdl_height,
start_x:start_x + mdl_width, :]
img2_tile = img2[start_y:start_y + mdl_height,
start_x:start_x + mdl_width, :]
edge_tile = edge[start_y:start_y + mdl_height,