Track from a different query frame without recomputing features

[ricshaw]

Hi, is it possible to track points from different query frames (not the first frame) without re-running the feature_extractor DPTHead? I.e. can I do just one forward pass of the DPTHead to get the features, and then run the tracker from different query frames? I tried changing the query index in BaseTrackerPredictor:

# Sample/extract the features of the query points in the query frame
query_track_feat = sample_features4d(fmaps[:, idx], coords[:, 0])

I also tried shuffling the extracted features so that the first frame in the feature tensor is different. But the tracking results are not good...

[jytime]

Hi @ricshaw ,

Thanks for sharing. This is a known issue and has been reported by other users. In theory it should work but something seems to be wrong. I am debugging this issue and retraining the model.