search

cv-stuttgart / MS_RAFT_plus

[IJCV 2023 / ECCV RVC 2022 Winner] Inference code for "MS-RAFT+: High Resolution Multi-Scale RAFT", the 2022 Robust Vision Challenge winner in the category "Optical Flow"
Other
25 stars 3 forks source link

On-demand cost sampling, backward pass #3

Open starsong98 opened 2 months ago

starsong98 commented 2 months ago

Thank you for sharing your code.

The MS-RAFT+ paper says that you implemented a backwards pass for the AlternateCorrBlock. Is this part of the released code? Or was it not released, alongside the rest of the training code? From what I can tell, the implementation of alt_cuda_corr in the released code seems nearly identical to that released in the original RAFT.

Thank you!

az-ja commented 2 months ago

Hello,

We did not release that in this project yet, but you can find it in our CCMR method under alt_cuda_corr.

starsong98 commented 2 months ago

Thank you! I will be sure to check it out.

xiaosongding commented 2 months ago

Thanks for sharing your code. when runing evalution.py code, there showed" corr = alt_cuda_corr.forward(fmap1_i, fmap2_i, coords_i, r) AttributeError: module 'alt_cuda_corr' has no attribute 'forward'". how to deal with this issue

az-ja commented 2 months ago

@xiaosongding Hi. You should compile the cuda module with a proper gcc version. After the cuda module is compiled successfully, it will run. Please check the readme. If you have further comments, please open another issue. Because this is not exactly relevant to the previous asked question.

xiaosongding commented 2 months ago

thanks for your response, I'm using windwos system, how to deal with this issue