Reverse tracking implementation seems to differ from paper

[ziliHarvey]

Thanks for the details on the forward and reverse order tracking fusion part mentioned in the paper!
However when I go through the code below

DetZero/tracking/detzero_track/models/tracking_modules/track_manager.py

Lines 109 to 142 in 051fad7

	# execute the reverse-order tracking stage
	if self.modules_dicts['reverse_tracking_config'].enable:
	frm_tracks = dict()
	reverse_tracks = list()
	keys = ['boxes_global', 'name', 'score', 'sample_idx',
	'hit', 'num_points', 'obj_ids']
	# transfer the tracking result to frame-level strutcure
	for tk_id in tk_result.keys():
	sample_idx = tk_result[tk_id]['sample_idx']
	for i, sa_idx in enumerate(sample_idx):
	if sa_idx not in frm_tracks.keys():
	frm_tracks[sa_idx] = defaultdict(list)
	frm_tracks[sa_idx]['start'].append(1 if i == 0 else 0)
	for key in keys:
	frm_tracks[sa_idx][key].append(tk_result[tk_id][key][i])
	for key, items in frm_tracks.items():
	for k, v in items.items():
	items[k] = np.array(v)
	for idx, frm_id in enumerate(frame_list[::-1]):
	frm_tk_data, reverse_tracks = self.reverse_tracking_module(
	frm_id, data_dict[frm_id], frm_tracks[frm_id], reverse_tracks
	)
	for key, val in frm_tk_data.items():
	for sub_key, sub_val in val.items():
	tk_result[key][sub_key] = np.insert(
	tk_result[key][sub_key], 0, sub_val, axis=0
	)
	tk_result[key]['pose'] = np.insert(
	tk_result[key]['pose'], 0, data_dict[frm_id]['pose'], axis=0
	)

I didn't see "These tracks are then associated together by a locationaware similarity matching score. Finally, we fuse the paired tracks with the WBF [42] strategy to further ameliorate the missing boxes and stable the motion state". I'm expecting something like a function to associate 2 trajectories (forward and backward for the same object) by bipartite matching and then another function to merge and refine the paired 2 trajectories into 1 complete trajectory.

Did I miss anything in the code?