Wei Jiang
I also think it should be `p1,p2 = source_pyr[k],target_pyr[k]`
Some tables: With random keypoints, we found cycle consistency is quite useful to obtain accurate matches. We rerun the experiments with FasterSparseEngine on ETH3D, and the performance slightly dropped. 
1. Evaluation code: Evaluation for ETH3D and KITTI should be straightforward, there should be one to one GT correspondences between image pairs. For HPatches, you need to correctly handle the...
Hi Vikram, The raw output of the network is [0, 1], and the input image tuple is at 256x256 resolution. In this case, [0, 1] == [0, height] in target...
Hi Vikram, Yes, your understanding about `cotr_corr_multiscale` is correct. `cotr_corr_base` uses 256x256 resolution image to provide the initial correspondences for the image pair, maybe there are too many outliers? `cotr_corr_base`...
hmmmm, I can see on the left the correspondences are from top to bottom, but there are a lot of horizontal correspondences which seems to be not making sense to...
if only one image is given, no, COTR can't predict the depth. If 2 images are given, and you know the intrinsics and extrinsics, then yes, you can triangulate the...
Hi Cheng, unfortunately, yes, it is supposed to be slow even on GPU...
Do you solve this issue by pip install or apt-get install?
Thanks! I’ll add this to the env file