def find_gt_bboxes(id, data_folder, segment_name, start_frame, end_frame):
""" This function is for finding the gt bboxes
Args:
id (str): id of the tracklet
gt_folder (str): root for data storage
segment_name (str): the segment to look at
start_frame (int): which frame to search
Return
list of bboxes
"""
gt_info = np.load(os.path.join(data_folder, 'gt_info',
'{:}.npz'.format(segment_name)),
allow_pickle=True)
ego_info = np.load(os.path.join(data_folder, 'ego_info',
'{:}.npz'.format(segment_name)),
allow_pickle=True)
bboxes, ids = gt_info['bboxes'][start_frame], gt_info['ids'][start_frame]
index = ids.index(id)
gts = list()
for i in range(start_frame + 1, end_frame + 1):
# first frame needs no evaluation
# so begin from start_frame + 1
frame_bboxes = gt_info['bboxes'][i]
frame_ids = gt_info['ids'][i]
index = frame_ids.index(id)
bbox = frame_bboxes[index]
bbox = BBox.array2bbox(bbox)
ego_matrix = ego_info[str(i)]
bbox = BBox.bbox2world(ego_matrix, bbox)
gts.append(bbox)
return gts
def find_bboxes(id, data_folder, segment_name, start_frame, end_frame):
""" In the SOT, the beginning frame is a GT BBox
This function is for finding the gt bbox
Args:
id (str): id of the tracklet
data_folder (str): root for data storage
segment_name (str): the segment to look at
start_frame (int): which frame to search
Return
BBox (numpy array): [x, y, z, h, l, w, h]
"""
gt_info = np.load(os.path.join(data_folder, 'gt_info',
'{:}.npz'.format(segment_name)),
allow_pickle=True)
ego_info = np.load(os.path.join(data_folder, 'ego_info',
'{:}.npz'.format(segment_name)),
allow_pickle=True)
bboxes, ids = gt_info['bboxes'][start_frame], gt_info['ids'][start_frame]
index = ids.index(id)
start_bbox = bboxes[index]
gts = list()
for i in range(start_frame, end_frame + 1):
frame_bboxes = gt_info['bboxes'][i]
frame_ids = gt_info['ids'][i]
index = frame_ids.index(id)
bbox = frame_bboxes[index]
bbox = BBox.array2bbox(bbox)
ego_matrix = ego_info[str(i)]
bbox = BBox.bbox2world(ego_matrix, bbox)
gts.append(bbox)
return start_bbox, gts
def preprocess(self):
""" Some data need eplicit transformation to the world coordinate:
point cloud, detection bboxes
"""
self.pc = utils.pc2world(self.ego, self.pc)
if self.dets is not None:
self.dets = [BBox.bbox2world(self.ego, det) for det in self.dets]
return
def tracker_api(configs,
id,
start_bbox,
start_frame,
data_loader,
track_len,
gts=None,
visualize=False):
""" api for the tracker
Args:
configs: model configuration read from config.yaml
id (str): each tracklet has an id
start_bbox ([x, y, z, yaw, l, w, h]): the beginning location of this id
data_loader (an iterator): iterator returning data of each incoming frame
Return:
{
frame_number0: pred_bbox0,
frame_number1: pred_bbox1,
...
frame_numberN: pred_bboxN
}
"""
tracker = sot_3d.Tracker(id=id,
configs=configs,
start_bbox=start_bbox,
start_frame=start_frame,
track_len=track_len)
tracklet_result = dict()
for frame_index in range(track_len):
print('////////////////////////////////////////')
print('Processing {:} {:} / {:}'.format(id, frame_index + 1,
track_len))
# initialize a tracker
frame_data = next(data_loader)
# if the first frame, add the start_bbox
input_bbox = None
if frame_index == 0:
input_bbox = BBox.bbox2world(frame_data['ego'],
BBox.array2bbox(start_bbox))
input_data = sot_3d.FrameData(ego_info=frame_data['ego'],
pc=frame_data['pc'],
start_bbox=input_bbox,
terrain=frame_data['terrain'],
dets=frame_data['dets'])
# run the frame level tracking
frame_output = tracker.track(input_data)
# the frame 0 may produce no output
if not frame_output:
continue
# if gt is not None, we may compare our prediction with gt
frame_result = compare_to_gt(frame_index, frame_output, gts)
max_frame_key = max(list(frame_result.keys()))
for i in range(max_frame_key + 1):
print('BBox0 : {:}'.format(frame_result[i]['bbox0']))
print('BBox1 : {:}'.format(frame_result[i]['bbox1']))
print('Motion : {:}'.format(frame_result[i]['motion']))
if gts:
print('GT BBox0 : {:}'.format(frame_result[i]['gt_bbox0']))
print('GT BBox1 : {:}'.format(frame_result[i]['gt_bbox1']))
print('GT Motion: {:}'.format(frame_result[i]['gt_motion']))
print('IOUS : {:} {:}'.format(frame_result[i]['iou2d'],
frame_result[i]['iou3d']))
print('\n')
tracklet_result[frame_index +
start_frame] = frame_result[max_frame_key]
if visualize:
frame_result_visualization(frame_result[max_frame_key],
tracker.input_data.pc)
return tracklet_result