search

autonomousvision / unimatch

[TPAMI'23] Unifying Flow, Stereo and Depth Estimation
https://haofeixu.github.io/unimatch/
MIT License
1.05k stars 106 forks source link

about multi-level feature extraction #20

Closed Steven-Xiong closed 1 year ago

Steven-Xiong commented 1 year ago

Hi, thanks for your great work. Is there possible ways to extract multi-level features in the optical flow estimation task? I notice there is a hyperparameter ' num_scales', but when I try to set it to 4, it seems it has correlation with the hyperparameter 'attn_splits_list', 'corr_radius_list' and 'prop_radius_list'. Could you show me how to adjust these three parameters if I would like to extract the 4 level upsample features? Thank you!

haofeixu commented 1 year ago

Hi, the 'attn_splits_list', 'corr_radius_list' and 'prop_radius_list' are lists where each element controls the corresponding paprameter of each scale. Thus if you use 4 scales, the length of these there parameters should also be 4.