DarkLabel VS sort

There is still no point retraining your model with its own output tho. You will get better results annotating it yourself. You can make use of this annotation tool https://github.com/darkpgmr/DarkLabel to save annotation time. If your training images are frames in a video, by using this tool you can propagate your bounding boxes into future frames. The boxes will eventually drift tho, so you will have to reinitialise the boxes every now and then.

Similari is a Rust/Python framework aimed at building sophisticated tracking systems. With Similari, you can develop highly efficient parallelized SORT, DeepSORT, and other sophisticated multiple-object tracking engines.

One interesting thing to consider is about using a new model's output as training data is the following. If you use some object tracking algorithm like SORT, you might be able to find frames where the object was "tracked" even though the model didn't detect the object (some type of flickering). In this case, you're able to run the model on video and make it better by giving it these new bounding boxes based on "tracking". Does that make sense?

The model is an ST-GCN trained in a completely unsupervised manner, from a LOT of skeleton sequences. Pose estimation was performed with AlphaPose, and tracked using SORT.

Yes it is possible! It's an addition to object detection, so I suggest you look for "object detection tracking". You can find articles like this one which refers to Python libraries like this one.

congratz! that's a great project. since you have the detector operating on videos, you could add a simple tracking-by-detection method like SORT. https://github.com/abewley/sort#using-sort-in-your-own-project it's really easy to use, and creates temporally constistent tracks by associating detections in successive frames.