Quick question on inference speed of LightStereo

[Jm7752]

I have a few quick questions on the recently released LightStereo-S Model being released. Currently I tried to use the infer python script and timed the evaluation time between the model inference and got something like 0.5Seconds.

I understand that we have different GPU, you are using a Nvidia RTX 3090 and I am using a RTX 3070, but this timing seems almost some as 2d- MobileStereoNet which is also like 0.5seconds on my computer.

This seems different from the paper where lightstereo should be an order of magnitude faster, so I am wondering if there is some additional optimization for lightstereo for speed than the infer.py script, or if you think it is just a hardware thing.

https://github.com/XiandaGuo/OpenStereo/blob/v2/tools/infer.py

Additionally, It doesn't seem to function very well on grayscale images. Do you guys have a trick for improving the performance on that or is that outside of the Scope for Lightstereo.

KITTI grayscale image Stereo Matching

Infered Disparity

Other examples out of distribution Infred Disparity

Thanks

[zhcm]

@Jm7752 please check tools/measure.py. The model's parameter count, computational cost, and inference speed will be analyzed. Ensure there are no other tasks running on your GPU before starting.

[Jm7752]

Ok thanks will do