How to get bbox for occluded frames using motion params?

[danial880]

Hi, my tracker is performing well it keeps track of the object even if it is occluded for 10 to 15 frames. Now how can I get BBox for these occluded frames? I have seen motion parameters in runtime tracker. How can I make use of those params?

[HELLORPG]

MeMOTR can not locate the occluded target. Because during training, most datasets do not annotate the occluded target (except MOT17), thus we expect the classification confidence of these occluded targets to be 0 and do not supervise their bounding boxes.

The motion module you mentioned is a deprecated processing. In our MeMOTR, the reference point of an occluded target is its last seen position. To use this motion-based processing, we roughly estimate the linear trajectory of the disappeared object to predict its possible re-appearance position, which is used as the reference point for subsequent frames. However, this process did not bring a significant improvement. On DanceTrack, it only improves no more than 0.5 HOTA. In line with less is more, I did not use this module in our final version.

[HELLORPG]

I'm not sure whether this motion-based module is ready in this open-source version. Here are the docs for this module:

• motion_min_length: the occluded object should appear in at least motion_min_length frames.
• motion_max_length: we only accept the trajectory of at most motion_max_length frames to predict the position.

One more thing, motion_lambda decides how much we trust this simple linear trajectory estimation.

[danial880]

Thanks for the info. In demo notebook we are using RuntimeTracker class and there is no motion_lambda implemented there.

[HELLORPG]

RuntimeTracker directly receives the previous tracks for updating. And the reference point is included in the param tracks: List[TrackInstances]. Therefore, you should use motion_lambda to compute the reference point outside and then input it in the RuntimeTrack.update. Like here we did.