def _compute_tracks_video(frames, dets):
nframes = len(frames)
video_tracks = []
next_track_id = FIRST_TRACK_ID
for frame_id in range(nframes):
frame_tracks = []
# each element is (roidb entry, idx in the dets/original roidb
cur_boxes = dets['all_boxes'][1][frame_id]
cur_poses = dets['all_keyps'][1][frame_id]
if (frame_id == 0):
matches = -np.ones((cur_boxes.shape[0], ))
else:
cur_frame_data = frames[frame_id]
prev_boxes = dets['all_boxes'][1][frame_id - 1]
prev_poses = dets['all_keyps'][1][frame_id - 1]
# 0-index to remove the other index to the dets structure
prev_frame_data = frames[frame_id-1]
matches = _compute_matches(
prev_frame_data, cur_frame_data,
prev_boxes, cur_boxes, prev_poses, cur_poses,
cost_types=cfg.TRACKING.DISTANCE_METRICS,
cost_weights=cfg.TRACKING.DISTANCE_METRIC_WTS,
bipart_match_algo=cfg.TRACKING.BIPARTITE_MATCHING_ALGO)
prev_tracks = video_tracks[frame_id - 1] if frame_id > 0 else None
for m in matches:
if m == -1: # didn't match to any
frame_tracks.append(next_track_id)
next_track_id += 1
if next_track_id >= MAX_TRACK_IDS:
next_track_id %= MAX_TRACK_IDS
else:
frame_tracks.append(prev_tracks[m])
video_tracks.append(frame_tracks)
return video_tracks
def _compute_tracks_video_lstm(frames, dets, lstm_model):
nframes = len(frames)
video_tracks = []
next_track_id = FIRST_TRACK_ID
# track_lstms contain track_id: <lstm_hidden_layer>
track_lstms = {}
for frame_id in range(nframes):
frame_tracks = []
# each element is (roidb entry, idx in the dets/original roidb)
cur_boxes = dets['all_boxes'][1][frame_id]
cur_poses = dets['all_keyps'][1][frame_id]
cur_boxposes = lstm_track_utils.encode_box_poses(cur_boxes, cur_poses)
# Compute LSTM next matches
# Need to keep prev_track_ids to make sure of ordering of output
prev_track_ids = video_tracks[frame_id - 1] if frame_id > 1 else []
match_scores = lstm_track_utils.compute_matching_scores(
track_lstms, prev_track_ids, cur_boxposes, lstm_model)
if match_scores.size > 0:
matches = _compute_matches(None,
None,
None,
None,
None,
None,
None,
None,
cfg.TRACKING.BIPARTITE_MATCHING_ALGO,
C=(-match_scores))
else:
matches = -np.ones((cur_boxes.shape[0], ))
prev_tracks = video_tracks[frame_id - 1] if frame_id > 0 else None
for m in matches:
if m == -1: # didn't match to any
frame_tracks.append(next_track_id)
next_track_id += 1
if next_track_id >= MAX_TRACK_IDS:
logger.warning(
'Exceeded max track ids ({})'.format(MAX_TRACK_IDS))
next_track_id %= MAX_TRACK_IDS
else:
frame_tracks.append(prev_tracks[m])
# based on the matches, update the lstm hidden weights
# Whatever don't get matched, start a new track ID. Whatever previous
# track IDs don't get matched, have to be deleted.
lstm_track_utils.update_lstms(track_lstms, prev_track_ids,
frame_tracks, cur_boxposes, lstm_model)
video_tracks.append(frame_tracks)
return video_tracks