pred_boxes_right = pred_boxes_right.squeeze()
pred_kpts = torch.cat(
(pred_kpts, kpts_type, max_prob, pred_left, pred_right), 2)
pred_kpts = pred_kpts.squeeze()
dim_orien = dim_orien.squeeze()
det_toc = time.time()
detect_time = det_toc - det_tic
img_path = imdb.img_left_path_at(i)
split_path = img_path.split('/')
image_number = split_path[len(split_path) - 1].split('.')[0]
calib_path = img_path.replace("image_2", "calib")
calib_path = calib_path.replace("png", "txt")
calib = kitti_utils.read_obj_calibration(calib_path)
label_path = calib_path.replace("calib", "label_2")
lidar_path = calib_path.replace("calib", "velodyne")
lidar_path = lidar_path.replace("txt", "bin")
im2show_left = np.copy(cv2.imread(imdb.img_left_path_at(i)))
im2show_right = np.copy(cv2.imread(imdb.img_right_path_at(i)))
pointcloud = kitti_utils.get_point_cloud(lidar_path, calib)
im_box = vis_utils.vis_lidar_in_bev(pointcloud,
width=im2show_left.shape[0] *
2)
for j in xrange(1, imdb.num_classes):
inds = torch.nonzero(scores[:, j] > eval_thresh).view(-1)
# if there is det
pred_left /= im_info[0, 2].data
pred_right /= im_info[0, 2].data
scores = scores.squeeze()
pred_boxes_left = pred_boxes_left.squeeze()
pred_boxes_right = pred_boxes_right.squeeze()
pred_kpts = torch.cat(
(pred_kpts, kpts_type, max_prob, pred_left, pred_right), 2)
pred_kpts = pred_kpts.squeeze()
dim_orien = dim_orien.squeeze()
det_toc = time.time()
detect_time = det_toc - det_tic
calib = kitti_utils.read_obj_calibration('demo/calib.txt')
im2show_left = np.copy(cv2.imread(img_l_path))
im2show_right = np.copy(cv2.imread(img_r_path))
pointcloud = kitti_utils.get_point_cloud('demo/lidar.bin', calib)
im_box = vis_utils.vis_lidar_in_bev(pointcloud,
width=im2show_left.shape[0] * 2)
for j in xrange(1, len(kitti_classes)):
inds = torch.nonzero(scores[:, j] > eval_thresh).view(-1)
# if there is det
if inds.numel() > 0:
cls_scores = scores[:, j][inds]
_, order = torch.sort(cls_scores, 0, True)
def _load_kitti_annotation(self, index):
if self._image_set == 'test':
objects = []
else:
filename = os.path.join(self._data_path, 'training', 'label_2',
index + '.txt')
calib_file = os.path.join(self._data_path, 'training', 'calib',
index + '.txt')
calib_it = kitti_utils.read_obj_calibration(calib_file)
im_left = cv2.imread(self.img_left_path_from_index(index))
objects_origin = kitti_utils.read_obj_data(filename, calib_it,
im_left.shape)
objects = []
objects_origin = self.remove_occluded_keypoints(objects_origin)
objects_origin = self.remove_occluded_keypoints(objects_origin,
left=False)
for i in range(len(objects_origin)):
if objects_origin[i].truncate < 0.98 and objects_origin[i].occlusion < 3 and \
(objects_origin[i].boxes[0].box[3] - objects_origin[i].boxes[0].box[1])>10 and \
objects_origin[i].cls in self._classes and \
objects_origin[i].boxes[0].visible_right - objects_origin[i].boxes[0].visible_left > 3 and\
objects_origin[i].boxes[1].visible_right - objects_origin[i].boxes[1].visible_left > 3:
objects.append(objects_origin[i])
f = calib_it.p2[0, 0]
cx = calib_it.p2[0, 2]
base_line = (calib_it.p2[0, 3] - calib_it.p3[0, 3]) / f
num_objs = len(objects)
boxes_left = np.zeros((num_objs, 4), dtype=np.float32)
boxes_right = np.zeros((num_objs, 4), dtype=np.float32)
boxes_merge = np.zeros((num_objs, 4), dtype=np.float32)
dim_orien = np.zeros((num_objs, 4), dtype=np.float32)
kpts = np.zeros((num_objs, 6), dtype=np.float32)
kpts_right = np.zeros((num_objs, 6), dtype=np.float32)
truncation = np.zeros((num_objs), dtype=np.float32)
occlusion = np.zeros((num_objs), dtype=np.float32)
gt_classes = np.zeros((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
for i in range(len(objects)):
cls = self._class_to_ind[objects[i].cls]
boxes_left[i, :] = objects[i].boxes[0].box
boxes_right[i, :] = objects[i].boxes[1].box
boxes_merge[i, :] = objects[i].boxes[2].box
dim_orien[i, 0:3] = objects[i].dim
dim_orien[i, 3] = objects[i].alpha
kpts[i, :4] = objects[i].boxes[0].keypoints
kpts[i, 4] = objects[i].boxes[0].visible_left
kpts[i, 5] = objects[i].boxes[0].visible_right
kpts_right[i, :4] = objects[i].boxes[1].keypoints
kpts_right[i, 4] = objects[i].boxes[1].visible_left
kpts_right[i, 5] = objects[i].boxes[1].visible_right
occlusion[i] = objects[i].occlusion
truncation[i] = objects[i].truncate
gt_classes[i] = cls
overlaps[i, cls] = 1.0
overlaps = scipy.sparse.csr_matrix(overlaps)
gt_subclasses = np.zeros((num_objs), dtype=np.int32)
gt_subclasses_flipped = np.zeros((num_objs), dtype=np.int32)
subindexes = np.zeros((num_objs, self.num_classes), dtype=np.int32)
subindexes_flipped = np.zeros((num_objs, self.num_classes),
dtype=np.int32)
subindexes = scipy.sparse.csr_matrix(subindexes)
subindexes_flipped = scipy.sparse.csr_matrix(subindexes_flipped)
return {
'boxes_left': boxes_left,
'boxes_right': boxes_right,
'boxes_merge': boxes_merge,
'dim_orien': dim_orien,
'kpts': kpts,
'kpts_right': kpts_right,
'truncation': truncation,
'occlusion': occlusion,
'gt_classes': gt_classes,
'igt_subclasses': gt_subclasses,
'gt_subclasses_flipped': gt_subclasses_flipped,
'gt_overlaps': overlaps,
'gt_subindexes': subindexes,
'gt_subindexes_flipped': subindexes_flipped,
'flipped': False
}
