def get_next_batch(self, batch_idx):
batch_size, orig_im_size = self.batch_size, self.orig_im_size
if batch_idx == self.num_batch - 1:
batch_size = self.last_batch_size
elif batch_idx >= self.num_batch:
print('batch index larger than the total number of batches')
return
im_heights, im_widths, num_scale = self.im_heights, self.im_widths, self.num_scale
ou_height, ou_width = self.output_height, self.output_width
images = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
depths = [
np.zeros((batch_size, im_heights[i], im_widths[i], 1))
for i in range(num_scale)
]
flows = [
np.zeros((batch_size, im_heights[i], im_widths[i], 2))
for i in range(num_scale)
]
label = np.zeros((batch_size, ou_height, ou_width))
for n in range(batch_size):
data_idx = self.sample_index[batch_idx * self.batch_size + n]
image = np.array(Image.open(self.meta['image'][data_idx]))
image = image / 255.0
depth = flowlib.read_disp_png(self.meta['depth'][data_idx])
flow = flowlib.read_flow_png(self.meta['flow'][data_idx])
flow = flow[:, :, 0:2]
if flow.shape[0] < self.orig_im_size[0]:
pad_length = self.orig_im_size[0] - flow.shape[0]
flow = np.pad(flow, ((pad_length, 0), (0, 0), (0, 0)),
'constant')
seg = np.array(Image.open(self.meta['seg'][data_idx]))
if self.data_augment:
image, depth, flow, seg = self.data_augmentation(
image, depth, flow, seg)
for i in range(num_scale):
images[i][n, :, :, :] = cv2.resize(
image, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
depths[i][n, :, :,
0] = cv2.resize(depth, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][n, :, :, :] = cv2.resize(
flow, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
label[n, :, :] = cv2.resize(seg, (ou_width, ou_height),
interpolation=cv2.INTER_NEAREST)
sample = {
'images': images,
'depths': depths,
'flows': flows,
'seg': label
}
return sample
def get_one_sample(self, image_name, depth_name, flow_name, seg_name):
batch_size, orig_im_size = 1, self.orig_im_size
im_heights, im_widths, num_scale = self.im_heights, self.im_widths, self.num_scale
ou_height, ou_width = self.output_height, self.output_width
images = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
depths = [
np.zeros((batch_size, im_heights[i], im_widths[i], 1))
for i in range(num_scale)
]
flows = [
np.zeros((batch_size, im_heights[i], im_widths[i], 2))
for i in range(num_scale)
]
label = np.zeros((batch_size, ou_height, ou_width))
for n in range(batch_size):
image = np.array(Image.open(image_name))
image = image / 255.0
image = image[:768, :, :]
depth = flowlib.read_disp_png(depth_name)
if depth.shape[0] > 768:
depth = depth[:768, :]
flow = flowlib.read_flow_png(flow_name)
flow = flow[:, :, 0:2]
if flow.shape[0] > 768:
flow = flow[:768, :, :]
seg = np.array(Image.open(seg_name))
seg = seg[:768, :, :]
for i in range(num_scale):
images[i][0, :, :, :] = cv2.resize(
image, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
depths[i][0, :, :,
0] = cv2.resize(depth, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][0, :, :, :] = cv2.resize(
flow, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
label[0, :, :] = cv2.resize(seg, (ou_width, ou_height),
interpolation=cv2.INTER_NEAREST)
sample = {
'images': images,
'flows': flows,
'depths': depths,
'seg': label
}
return sample
def get_one_sample(self, image_name, depth_name, flow_name, box_name):
image = np.array(Image.open(image_name))
image = image / 255.0
depth = flowlib.read_disp_png(depth_name)
flow = flowlib.read_flow_png(flow_name)
box_and_label = []
with open(box_name) as txt_file:
box_info = txt_file.readlines()
for row in box_info:
row = row.strip().split(' ')
if not row[0] in ['Car']:
continue
row[0] = self.class_map[row[0]]
box_and_label.append(row)
box_and_label = np.array(box_and_label).astype(np.float)
if box_and_label.shape[0] == 0:
box = []
label = []
else:
box = box_and_label[:, 4:8]
label = box_and_label[:, 0]
batch_size = 1
orig_im_size = self.orig_im_size
im_heights, im_widths, num_scale = self.im_heights, self.im_widths, self.num_scale
ou_heights, ou_widths = self.output_heights, self.output_widths
images = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
orig_image = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 3))
depths = [
np.zeros((batch_size, im_heights[i], im_widths[i], 1))
for i in range(num_scale)
]
orig_depth = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 1))
flows = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
orig_flow = np.zeros((batch_size, orig_im_size[0], orig_im_size[1], 3))
boxes = []
label_maps = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 1))
for i in range(num_scale)
]
offsets = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 4))
for i in range(num_scale)
]
# image, depth, flow, box = self.crop_image(image, depth, flow, box)
im_height, im_width = image.shape[0], image.shape[1]
for i in range(num_scale):
images[i][0, :, :, :] = cv2.resize(image,
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
depths[i][0, :, :, 0] = cv2.resize(depth,
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][0, :, :, 0:2] = cv2.resize(flow[:, :, 0:2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][0, :, :, 2] = cv2.resize(flow[:, :, 2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_NEAREST)
orig_image[0, :, :, :] = cv2.resize(
image, (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_depth[0, :, :, 0] = cv2.resize(
depth, (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[0, :, :, 0:2] = cv2.resize(
flow[:, :, 0:2], (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[0, :, :,
2] = cv2.resize(flow[:, :, 2],
(self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_NEAREST)
if len(box) > 0:
box[:, 0] = box[:, 0] * 1.0 / im_width
box[:, 1] = box[:, 1] * 1.0 / im_height
box[:, 2] = box[:, 2] * 1.0 / im_width
box[:, 3] = box[:, 3] * 1.0 / im_height
boxes.append(box)
x = (box[:, 0] + box[:, 2]) * 1.0 / 2
y = (box[:, 1] + box[:, 3]) * 1.0 / 2
w = (box[:, 2] - box[:, 0]) * 1.0
h = (box[:, 3] - box[:, 1]) * 1.0
for k in range(box.shape[0]):
for i in range(num_scale):
# if w[k] < 1.0 / ou_widths[i] or h[k] < 1.0 / ou_heights[i]:
# continue
# x_c = np.int(np.round(x[k] * ou_widths[i]))
# y_c = np.int(np.round(y[k] * ou_heights[i]))
# label_maps[i][0, y_c, x_c, 0] = 1 # Only works for car detection
# # x1 = max(x_c - 1, 0)
# # x2 = min(x_c + 1, self.output_widths[i]-1)
# # y1 = max(y_c - 1, 0)
# # y2 = min(y_c + 1, self.output_heights[i]-1)
# # label_maps[i][0, y1:y2, x1:x2, 0] = 1 # Only works for car detection
# # offsets[i][0, y_c, x_c, 0] = w[k]
# # offsets[i][0, y_c, x_c, 1] = h[k]
# offsets[i][0, y_c, x_c, 0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
# offsets[i][0, y_c, x_c, 1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
# offsets[i][0, y_c, x_c, 2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
# offsets[i][0, y_c, x_c, 3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
if w[k] < 2.0 / ou_widths[i] or h[k] < 2.0 / ou_heights[i]:
continue
x_c = np.int(np.floor(x[k] * ou_widths[i]))
y_c = np.int(np.floor(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][0, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][0, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][0, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.floor(x[k] * ou_widths[i]))
y_c = np.int(np.ceil(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][0, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][0, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][0, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.ceil(x[k] * ou_widths[i]))
y_c = np.int(np.floor(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][0, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][0, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][0, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.ceil(x[k] * ou_widths[i]))
y_c = np.int(np.ceil(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][0, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][0, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][0, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][0, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
else:
boxes.append([])
labels = label
for i in range(num_scale):
images[i] = images[i].transpose((0, 3, 1, 2))
depths[i] = depths[i].transpose((0, 3, 1, 2))
flows[i] = flows[i].transpose((0, 3, 1, 2))
label_maps[i] = label_maps[i].transpose((0, 3, 1, 2))
offsets[i] = offsets[i].transpose((0, 3, 1, 2))
orig_image = orig_image.transpose((0, 3, 1, 2))
orig_depth = orig_depth.transpose((0, 3, 1, 2))
orig_flow = orig_flow.transpose((0, 3, 1, 2))
return images, orig_image, depths, orig_depth, flows, orig_flow, boxes, \
label_maps, offsets
def get_next_batch(self, status='train', cnt=0, index=None):
if status == 'train':
anno = self.train_anno
elif status == 'test':
anno = self.test_anno
else:
logging.error('Error: wrong status')
if index is None:
index = np.arange(len(anno['img']))
batch_size, orig_im_size = self.batch_size, self.orig_im_size
im_heights, im_widths, num_scale = self.im_heights, self.im_widths, self.num_scale
ou_heights, ou_widths = self.output_heights, self.output_widths
images = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
orig_image = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 3))
depths = [
np.zeros((batch_size, im_heights[i], im_widths[i], 1))
for i in range(num_scale)
]
orig_depth = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 1))
flows = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
orig_flow = np.zeros((batch_size, orig_im_size[0], orig_im_size[1], 3))
boxes = []
label_maps = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 1))
for i in range(num_scale)
]
offsets = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 4))
for i in range(num_scale)
]
restart = False
for n in range(batch_size):
if cnt >= len(index):
cnt = 0
restart = True
image = np.array(Image.open(anno['img'][index[cnt]]))
image = image / 255.0
depth = flowlib.read_disp_png(anno['depth'][index[cnt]])
flow = flowlib.read_flow_png(anno['flow'][index[cnt]])
box = np.array(anno['box'][index[cnt]])
label = anno['label'][index[cnt]]
if status == 'train':
image, depth, flow, box = self.data_augmentation(
image, depth, flow, box)
image, depth, flow, box = self.crop_image(
image, depth, flow, box)
im_height, im_width = image.shape[0], image.shape[1]
for i in range(num_scale):
images[i][n, :, :, :] = cv2.resize(
image, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
depths[i][n, :, :,
0] = cv2.resize(depth, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][n, :, :,
0:2] = cv2.resize(flow[:, :, 0:2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][n, :, :,
2] = cv2.resize(flow[:, :, 2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_NEAREST)
if status == 'train':
orig_image[n, :, :, :] = image
orig_depth[n, :, :, 0] = depth
orig_flow[n, :, :, :] = flow
else:
orig_image[n, :, :, :] = cv2.resize(
image, (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_depth[n, :, :, 0] = cv2.resize(
depth, (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[n, :, :, 0:2] = cv2.resize(
flow[:, :,
0:2], (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[n, :, :, 2] = cv2.resize(
flow[:, :,
2], (self.orig_im_size[1], self.orig_im_size[0]),
interpolation=cv2.INTER_NEAREST)
if len(box) > 0:
box[:, 0] = box[:, 0] * 1.0 / im_width
box[:, 1] = box[:, 1] * 1.0 / im_height
box[:, 2] = box[:, 2] * 1.0 / im_width
box[:, 3] = box[:, 3] * 1.0 / im_height
boxes.append(box)
x = (box[:, 0] + box[:, 2]) * 1.0 / 2
y = (box[:, 1] + box[:, 3]) * 1.0 / 2
w = (box[:, 2] - box[:, 0]) * 1.0
h = (box[:, 3] - box[:, 1]) * 1.0
for k in range(box.shape[0]):
for i in range(num_scale):
# if w[k] < 1.0 / ou_widths[i] or h[k] < 1.0 / ou_heights[i]:
# continue
# x_c = np.int(np.round(x[k] * ou_widths[i]))
# y_c = np.int(np.round(y[k] * ou_heights[i]))
# if x_c < 0 or x_c >= self.output_widths[i] or y_c < 0 or y_c >= self.output_heights[i]:
# continue
# label_maps[i][n, y_c, x_c, 0] = 1 # Only works for car detection
# # x1 = max(x_c - 1, 0)
# # x2 = min(x_c + 1, self.output_widths[i]-1)
# # y1 = max(y_c - 1, 0)
# # y2 = min(y_c + 1, self.output_heights[i]-1)
# # label_maps[i][n, y1:y2, x1:x2, 0] = 1 # Only works for car detection
# offsets[i][n, y_c, x_c, 0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
# offsets[i][n, y_c, x_c, 1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
# offsets[i][n, y_c, x_c, 2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
# offsets[i][n, y_c, x_c, 3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
if w[k] < 2.0 / ou_widths[i] or h[
k] < 2.0 / ou_heights[i]:
continue
x_c = np.int(np.floor(x[k] * ou_widths[i]))
y_c = np.int(np.floor(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][n, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][n, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][n, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.floor(x[k] * ou_widths[i]))
y_c = np.int(np.ceil(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][n, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][n, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][n, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.ceil(x[k] * ou_widths[i]))
y_c = np.int(np.floor(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][n, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][n, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][n, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
x_c = np.int(np.ceil(x[k] * ou_widths[i]))
y_c = np.int(np.ceil(y[k] * ou_heights[i]))
if x_c < 0 or x_c >= self.output_widths[
i] or y_c < 0 or y_c >= self.output_heights[i]:
continue
label_maps[i][n, y_c, x_c,
0] = 1 # Only works for car detection
offsets[i][n, y_c, x_c,
0] = box[k, 0] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
1] = box[k, 1] - y_c * 1.0 / ou_heights[i]
offsets[i][n, y_c, x_c,
2] = box[k, 2] - x_c * 1.0 / ou_widths[i]
offsets[i][n, y_c, x_c,
3] = box[k, 3] - y_c * 1.0 / ou_heights[i]
else:
boxes.append(np.zeros((0, 4)))
cnt = cnt + 1
for i in range(num_scale):
images[i] = images[i].transpose((0, 3, 1, 2))
depths[i] = depths[i].transpose((0, 3, 1, 2))
flows[i] = flows[i].transpose((0, 3, 1, 2))
label_maps[i] = label_maps[i].transpose((0, 3, 1, 2))
offsets[i] = offsets[i].transpose((0, 3, 1, 2))
orig_image = orig_image.transpose((0, 3, 1, 2))
orig_depth = orig_depth.transpose((0, 3, 1, 2))
orig_flow = orig_flow.transpose((0, 3, 1, 2))
return images, orig_image, depths, orig_depth, flows, orig_flow, boxes, \
label_maps, offsets, cnt, restart
def get_one_sample(self, image_name, depth_name, flow_name, box_name):
batch_size, orig_im_size = 1, self.orig_im_size
im_heights, im_widths, num_scale = self.im_heights, self.im_widths, self.num_scale
ou_heights, ou_widths = self.output_heights, self.output_widths
images = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
depths = [
np.zeros((batch_size, im_heights[i], im_widths[i], 1))
for i in range(num_scale)
]
flows = [
np.zeros((batch_size, im_heights[i], im_widths[i], 3))
for i in range(num_scale)
]
heatmaps = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 1))
for i in range(num_scale)
]
offsets = [
np.zeros((batch_size, ou_heights[i], ou_widths[i], 4))
for i in range(num_scale)
]
orig_image = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 3))
orig_depth = np.zeros(
(batch_size, orig_im_size[0], orig_im_size[1], 1))
orig_flow = np.zeros((batch_size, orig_im_size[0], orig_im_size[1], 3))
boxes = []
for n in range(batch_size):
image = np.array(Image.open(image_name))
image = image / 255.0
depth = flowlib.read_disp_png(depth_name)
flow = flowlib.read_flow_png(flow_name)
box = read_box(box_name)
if self.data_augment:
image, depth, flow, box = self.flip_image(
image, depth, flow, box)
image, depth, flow, box = self.resize_crop_image(
image, depth, flow, box)
for i in range(num_scale):
images[i][n, :, :, :] = cv2.resize(
image, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
depths[i][n, :, :,
0] = cv2.resize(depth, (im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][n, :, :,
0:2] = cv2.resize(flow[:, :, 0:2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_AREA)
flows[i][n, :, :,
2] = cv2.resize(flow[:, :, 2],
(im_widths[i], im_heights[i]),
interpolation=cv2.INTER_NEAREST)
im_height, im_width = image.shape[0], image.shape[1]
box = self.rescale_box(box, im_height, im_width)
boxes.append(box)
xb, yb, wb, hb = self.get_box_center_size(box)
for k in range(box.shape[0]):
for i in range(num_scale):
if wb[k] < 2.0 / ou_widths[i] or hb[
k] < 2.0 / ou_heights[i]:
continue
x = np.int(np.floor(xb[k] * ou_widths[i]))
y = np.int(np.floor(yb[k] * ou_heights[i]))
if x < 0 or x >= ou_widths[i] or y < 0 or y >= ou_heights[
i]:
continue
heatmaps[i][n, y, x, 0] = 1 # Only works for car detection
offsets[i][n, y, x, 0] = box[k, 0] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
1] = box[k, 1] - y * 1.0 / ou_heights[i]
offsets[i][n, y, x, 2] = box[k, 2] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
3] = box[k, 3] - y * 1.0 / ou_heights[i]
x = np.int(np.floor(xb[k] * ou_widths[i]))
y = np.int(np.ceil(yb[k] * ou_heights[i]))
if x < 0 or x >= ou_widths[i] or y < 0 or y >= ou_heights[
i]:
continue
heatmaps[i][n, y, x, 0] = 1 # Only works for car detection
offsets[i][n, y, x, 0] = box[k, 0] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
1] = box[k, 1] - y * 1.0 / ou_heights[i]
offsets[i][n, y, x, 2] = box[k, 2] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
3] = box[k, 3] - y * 1.0 / ou_heights[i]
x = np.int(np.ceil(xb[k] * ou_widths[i]))
y = np.int(np.floor(yb[k] * ou_heights[i]))
if x < 0 or x >= ou_widths[i] or y < 0 or y >= ou_heights[
i]:
continue
heatmaps[i][n, y, x, 0] = 1 # Only works for car detection
offsets[i][n, y, x, 0] = box[k, 0] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
1] = box[k, 1] - y * 1.0 / ou_heights[i]
offsets[i][n, y, x, 2] = box[k, 2] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
3] = box[k, 3] - y * 1.0 / ou_heights[i]
x = np.int(np.ceil(xb[k] * ou_widths[i]))
y = np.int(np.ceil(yb[k] * ou_heights[i]))
if x < 0 or x >= ou_widths[i] or y < 0 or y >= ou_heights[
i]:
continue
heatmaps[i][n, y, x, 0] = 1 # Only works for car detection
offsets[i][n, y, x, 0] = box[k, 0] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
1] = box[k, 1] - y * 1.0 / ou_heights[i]
offsets[i][n, y, x, 2] = box[k, 2] - x * 1.0 / ou_widths[i]
offsets[i][n, y, x,
3] = box[k, 3] - y * 1.0 / ou_heights[i]
if self.data_augment:
orig_image[n, :, :, :] = image
orig_depth[n, :, :, 0] = depth
orig_flow[n, :, :, :] = flow
else:
orig_image[n, :, :, :] = cv2.resize(
image, (orig_im_size[1], orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_depth[n, :, :,
0] = cv2.resize(depth,
(orig_im_size[1], orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[n, :, :,
0:2] = cv2.resize(flow[:, :, 0:2],
(orig_im_size[1], orig_im_size[0]),
interpolation=cv2.INTER_AREA)
orig_flow[n, :, :,
2] = cv2.resize(flow[:, :, 2],
(orig_im_size[1], orig_im_size[0]),
interpolation=cv2.INTER_NEAREST)
sample = {
'images': images,
'flows': flows,
'depths': depths,
'heatmaps': heatmaps,
'offsets': offsets,
'orig_image': orig_image,
'orig_depth': orig_depth,
'orig_flow': orig_flow,
'boxes': boxes
}
return sample