search

XuelianCheng / LEAStereo

Hierarchical Neural Architecture Searchfor Deep Stereo Matching (NeurIPS 2020)
MIT License
256 stars 51 forks source link

Reproduce results on the SceneFlow dataset #23

Closed ShichenLiu closed 3 years ago

ShichenLiu commented 3 years ago

Hi,

I am wondering how to reproduce the sceneflow results. I followed the author's scripts that train the architecture searched for 20 epochs with a fixed learning rate of 0.001. The results were far from the reported numbers in the paper (1.01EPE against 0.78EPE). So I would like to know any learning rate scheduling that could reproduce the sceneflow results (also any checkpoint available? It seems the sceneflow checkpoint is missing)

Thanks!

alekseynp commented 3 years ago

https://github.com/XuelianCheng/LEAStereo/blob/master/run/sceneflow/best/checkpoint/best.pth

alekseynp commented 3 years ago

@ShichenLiu were you able to reproduce their results?

ShichenLiu commented 3 years ago

Hi @alekseynp No I am not. Also I tried their kitti15 code, and the val error was far from their performance on the leaderboard (D1-all 2.3% vs 1.6%). I am not sure if there is any additional training steps or arguments.

gautamsreekumar commented 3 years ago

Hi @ShichenLiu, could you please tell me what all changes you made to the original code to get 2.3%? I am getting 6.47% validation error with the default parameters.