which detection is assigned to a given track_id and which one are not

[fenaux]

Search before asking

• I have searched the Yolo Tracking issues and found no similar bug report.

Question

Thanks for this repo

In one frame I have 10 boxes for persons lets say they have corresponding indexes 0 for first box, 1 for second ....9 for the 10th.
I only have 7 tracks

Is there a way to know directly
the index of the box that is assigned to a given track_id
the indexes of the boxes that are not assigned to any track (could be track "-1" like values not in a cluster for sckilt learn dbscan)

Thanks for your help

[mikel-brostrom]

The index given to each bbox depends on the association rounds of each tracking algorithm. The assigned IDs are generated in ascending order. When a track ID is lost it can be recovered for a limited period of time, in order not to flood the memory with lost tracks, otherwise it is lost forever.

[fenaux]

Thanks for answering me.
But I must admit I could not figure how to find which box is assigned to which track from it

So here is the way I do it now

while True:
    ret, im = vid.read()
    if not ret: break

    # substitute by your object detector, output has to be N X (x, y, x, y, conf, cls)
    in_this_frame = np.where(in_frame == i_frame)[0]
    dets = boxes[in_this_frame]

    tracker.update(dets, im) # --> M X (x, y, x, y, id, conf, cls, ind)
    tracker.plot_results(im, show_trajectories=True)

    new_boxes = []
    new_ids = []
    for a in tracker.active_tracks:
                if a.history_observations:
                    if len(a.history_observations) > 2:
                        box = a.history_observations[-1]
                        new_boxes.append(box)
                        new_ids.append(a.id)
                        new_obs = np.hstack((i_frame,box, a.id))
                        if len(all_tracks) == 0: all_tracks = new_obs.copy()
                        else: all_tracks = np.row_stack((all_tracks, new_obs))

    video_writer.write(im)

    pairdist = distance.cdist(dets, new_boxes, 'euclidean')
    pairs = np.where(pairdist == 0)  # this line give the assignment  between dets and new_ids
 
    i_frame += 1

With DeepOcSort
And other thing surprised me, quiet often, they are more boxes in new_boxes (i.e. boxes corresponding to active tracks) than boxes in dets in my case on average 10% more.
And is clear in the tracker video that some boxes are false positive (aka ids 29 and 76 in image). https://drive.google.com/file/d/1hBBVOYyZkLEaNC348jbI-NDNi6QCCul7/view?usp=sharing
An option to plot boxes only corresponding to "dets" would make sense. May be det_id is what I am looking for ?

With BotSort
Behavior is different in my case I always got len(new_boxes) < len(dets)
and all new_boxes are close to on det