search

AssafSinger94 / dino-tracker

Official Pytorch Implementation for “DINO-Tracker: Taming DINO for Self-Supervised Point Tracking in a Single Video”
MIT License
361 stars 39 forks source link

About correlation part #1

Closed TomTomTommi closed 5 months ago

TomTomTommi commented 5 months ago

The corr_maps calculated in the code are in the embedding dimensions, right?

Is it possible to reshape the embeddings to normal feature sizes (B, C, H', W') and then build the cost volume?

tnarek commented 5 months ago

hi @TomTomTommi , thanks for your question. Could you please point to the specific part in the code?

In Tracker.get_corr_maps_for_frame_set, we do reshape to your mentioned dimensions before computing the cost volume.