add support for mmdet tracking

[rolson24]

Description

Feature request from #1381

Add support for MMDetection trackers by modifying from_mmdetection to support importing MMDetection TrackDataSamples into supvervision Detections. This also allows for easy integration with the [MASA tracker] (https://github.com/siyuanliii/masa/tree/1d5d42cdd75276c9a8e918639907916f75a761da)

Type of change

• New feature (non-breaking change which adds functionality)

How has this change been tested, please provide a testcase or example of how you tested the change?

Tested with this colab notebook

Any specific deployment considerations

Right now the pre-trained MMDetection models are performing quite poorly for tracking, and the built in config files need to be fixed, but if you train your own detection model and write your own config, they work. All you have to do is call `tracker = init_track_model(TRACKER_CONFIG, device=device)' and then for each frame call:

track_result = inference_mot(tracker, frame,
                        frame_id=frame_idx,
                        video_len=video_info.total_frames,
                        )

[onuralpszr]

@rolson24 is there chance tracker can be empty ? If yes we should do same thing we did for mask as well.

[SkalskiP]

@rolson24 I won't be able to test it today. I'll leave it to @LinasKo. He should be back on Monday.

[rolson24]

@rolson24 is there chance tracker can be empty ? If yes we should do same thing we did for mask as well.

I tested it and it seems like if there are no tracks, then pred_track_instances won't exist, so it will just use the pred_instances which are the detection results. But it probably wouldn't hurt to add it especially if there is a tracker that implements pred_track_instances slightly differently.

[Borda]

Thanks for working on this and the related issue #1381! Since we're deprecating the tracker module in supervision (see #2215) in favor of the new standalone trackers package, tracking integrations like this are best continued there. The from_mmdetection Detections support part may still be relevant separately — feel free to open a focused PR for that if needed. Thanks for your contribution! 🙏