search

ubc-vision / COTR

Code release for "COTR: Correspondence Transformer for Matching Across Images"(ICCV 2021)
Apache License 2.0
460 stars 58 forks source link

Optical flow estimation #45

Open Wyatt299 opened 1 year ago

Wyatt299 commented 1 year ago

Could you please provide a demo of optical flow estimation？Thank you.

jiangwei221 commented 1 year ago

Hey Wyatt, we obtain the optical flow/dense correspondence map by interpolating the sparse correspondences, as shown here: https://github.com/ubc-vision/COTR/blob/5c9363f1ca5c1825d999f875e8eb58914a64b01d/demo_single_pair.py#L42. The interpolated optical flow/dense correspondence has a smaller area than the GT, that's because the our filtering strategy will prune out some untrustworthy points around borders. As for how to select query points, 1. randomly select, 2. select points with GT flow(KITTI).

Wyatt299 commented 1 year ago

Great answer, thanks a lot, hope to post a demo on optical flow estimation later.