search

haofeixu / aanet

[CVPR'20] AANet: Adaptive Aggregation Network for Efficient Stereo Matching
Apache License 2.0
524 stars 102 forks source link

Disparity Refinement #5

Closed youmi-zym closed 4 years ago

youmi-zym commented 4 years ago

Thanks for your brilliant work. May I ask the performance of your AANet or AANet+ after removing the refinement module. Thanks. By the way, I know there is no implementation of 3D deformable convolution, have you thinking about implementing it and releasing 3D ISA and CSA?

haofeixu commented 4 years ago

Thanks for your interest. Unfortunately, we didn't have such a comparison. However, we anticipate the performance will be not that good, as the resolution of estimated disparity without refinement is relatively low and the residual learning (i.e., refinement) scheme has been proved to be effective in many previous works (e.g., StereoNet). You can train with our code to see what happens. Sorry about that we have no plan to extend the deformable conv to 3D currently.

youmi-zym commented 4 years ago

Thanks

XYZ-qiyh commented 4 years ago

Hello, @haofeixu As I know, the scale_factor is the ratio of left_img Height and low_disp Height. But, why the disp need to multiply the scale_factor?

https://github.com/haofeixu/aanet/blob/5e2359cfea9c38ebcddf5e750ec9ad1ac92725fc/nets/refinement.py#L45-L47

haofeixu commented 4 years ago

Hi @Todd-Qi ,

As the low resolution disparity is upsampled to higher resolution, it's magnitude needs to be scaled correspondingly. That is, a 1-pixel disparity at 1/4 resolution corresponds to 2-pixel for 1/2 resolution.

XYZ-qiyh commented 4 years ago

Thanks for your clear explanation.