def process_sample_data(nuscenes, map_data, sample_data, lidar, config):
# Render static road geometry masks
map_masks = nusc_utils.get_map_masks(nuscenes, map_data, sample_data,
config.map_extents,
config.map_resolution)
# Render dynamic object masks
obj_masks = nusc_utils.get_object_masks(nuscenes, sample_data,
config.map_extents,
config.map_resolution)
masks = np.concatenate([map_masks, obj_masks], axis=0)
# Ignore regions of the BEV which are outside the image
sensor = nuscenes.get('calibrated_sensor',
sample_data['calibrated_sensor_token'])
intrinsics = np.array(sensor['camera_intrinsic'])
masks[-1] |= ~get_visible_mask(intrinsics, sample_data['width'],
config.map_extents, config.map_resolution)
# Transform lidar points into camera coordinates
cam_transform = nusc_utils.get_sensor_transform(nuscenes, sample_data)
cam_points = transform(np.linalg.inv(cam_transform), lidar)
masks[-1] |= get_occlusion_mask(cam_points, config.map_extents,
config.map_resolution)
# Encode masks as integer bitmask
labels = encode_binary_labels(masks)
# Save outputs to disk
output_path = os.path.join(os.path.expandvars(config.label_root),
sample_data['token'] + '.png')
Image.fromarray(labels.astype(np.int32), mode='I').save(output_path)
def process_sample(nuscenes, map_data, sample, config):
# Load the lidar point cloud associated with this sample
lidar_data = nuscenes.get('sample_data', sample['data']['LIDAR_TOP'])
lidar_pcl = nusc_utils.load_point_cloud(nuscenes, lidar_data)
# Transform points into world coordinate system
lidar_transform = nusc_utils.get_sensor_transform(nuscenes, lidar_data)
lidar_pcl = transform(lidar_transform, lidar_pcl)
# Iterate over sample data
for camera in nusc_utils.CAMERA_NAMES:
sample_data = nuscenes.get('sample_data', sample['data'][camera])
process_sample_data(nuscenes, map_data, sample_data, lidar_pcl, config)