def occlusion(label, xyz):
if xyz.shape[0] == 0:
return 0
normals, lower, upper = dataset.box3d_to_normals(label)
projected = np.matmul(xyz, np.transpose(normals))
x_cover_rate = (np.max(projected[:, 0])-np.min(projected[:, 0]))\
/(upper[0] - lower[0])
y_cover_rate = (np.max(projected[:, 1])-np.min(projected[:, 1]))\
/(upper[1] - lower[1])
z_cover_rate = (np.max(projected[:, 2])-np.min(projected[:, 2]))\
/(upper[2] - lower[2])
return x_cover_rate * y_cover_rate * z_cover_rate
# setup model =================================================================
BOX_ENCODING_LEN = get_encoding_len(config['box_encoding_method'])
box_encoding_fn = get_box_encoding_fn(config['box_encoding_method'])
box_decoding_fn = get_box_decoding_fn(config['box_encoding_method'])
if config['input_features'] == 'irgb':
t_initial_vertex_features = tf.placeholder(dtype=tf.float32,
shape=[None, 4])
elif config['input_features'] == 'rgb':
t_initial_vertex_features = tf.placeholder(dtype=tf.float32,
shape=[None, 3])
elif config['input_features'] == '0000':
t_initial_vertex_features = tf.placeholder(dtype=tf.float32,
shape=[None, 4])
elif config['input_features'] == 'i000':
t_initial_vertex_features = tf.placeholder(dtype=tf.float32,
shape=[None, 4])
elif config['input_features'] == 'i':
def fetch_data(dataset, frame_idx, train_config, config):
aug_fn = preprocess.get_data_aug(train_config['data_aug_configs'])
BOX_ENCODING_LEN = get_encoding_len(config['box_encoding_method'])
box_encoding_fn = get_box_encoding_fn(config['box_encoding_method'])
box_decoding_fn = get_box_decoding_fn(config['box_encoding_method'])
graph_generate_fn = get_graph_generate_fn(config['graph_gen_method'])
cam_rgb_points = dataset.get_cam_points_in_image_with_rgb(
frame_idx, config['downsample_by_voxel_size'])
box_label_list = dataset.get_label(frame_idx)
if 'crop_aug' in train_config:
cam_rgb_points, box_label_list = sampler.crop_aug(
cam_rgb_points,
box_label_list,
sample_rate=train_config['crop_aug']['sample_rate'],
parser_kwargs=train_config['crop_aug']['parser_kwargs'])
cam_rgb_points, box_label_list = aug_fn(cam_rgb_points, box_label_list)
(vertex_coord_list, keypoint_indices_list, edges_list) = \
graph_generate_fn(cam_rgb_points.xyz, **config['graph_gen_kwargs'])
if config['input_features'] == 'irgb':
input_v = cam_rgb_points.attr
elif config['input_features'] == '0rgb':
input_v = np.hstack([
np.zeros((cam_rgb_points.attr.shape[0], 1)),
cam_rgb_points.attr[:, 1:]
])
elif config['input_features'] == '0000':
input_v = np.zeros_like(cam_rgb_points.attr)
elif config['input_features'] == 'i000':
input_v = np.hstack([
cam_rgb_points.attr[:, [0]],
np.zeros((cam_rgb_points.attr.shape[0], 3))
])
elif config['input_features'] == 'i':
input_v = cam_rgb_points.attr[:, [0]]
elif config['input_features'] == '0':
input_v = np.zeros((cam_rgb_points.attr.shape[0], 1))
last_layer_graph_level = config['model_kwargs']['layer_configs'][-1][
'graph_level']
last_layer_points_xyz = vertex_coord_list[last_layer_graph_level + 1]
if config['label_method'] == 'yaw':
cls_labels, boxes_3d, valid_boxes, label_map = \
dataset.assign_classaware_label_to_points(box_label_list,
last_layer_points_xyz,
expend_factor=train_config.get('expend_factor', (1.0, 1.0, 1.0)))
if config['label_method'] == 'Car':
cls_labels, boxes_3d, valid_boxes, label_map = \
dataset.assign_classaware_car_label_to_points(box_label_list,
last_layer_points_xyz,
expend_factor=train_config.get('expend_factor', (1.0, 1.0, 1.0)))
if config['label_method'] == 'Pedestrian_and_Cyclist':
(cls_labels, boxes_3d, valid_boxes, label_map) =\
dataset.assign_classaware_ped_and_cyc_label_to_points(
box_label_list, last_layer_points_xyz,
expend_factor=train_config.get('expend_factor', (1.0, 1.0, 1.0)))
encoded_boxes = box_encoding_fn(cls_labels, last_layer_points_xyz,
boxes_3d, label_map)
input_v = input_v.astype(np.float32)
vertex_coord_list = [p.astype(np.float32) for p in vertex_coord_list]
keypoint_indices_list = [e.astype(np.int32) for e in keypoint_indices_list]
edges_list = [e.astype(np.int32) for e in edges_list]
cls_labels = cls_labels.astype(np.int32)
encoded_boxes = encoded_boxes.astype(np.float32)
valid_boxes = valid_boxes.astype(np.float32)
return (input_v, vertex_coord_list, keypoint_indices_list, edges_list,
cls_labels, encoded_boxes, valid_boxes)