search

uber-research / DeepPruner

DeepPruner: Learning Efficient Stereo Matching via Differentiable PatchMatch (ICCV 2019)
Other
354 stars 41 forks source link

about the batchsize #13

Closed xubin1994 closed 4 years ago

xubin1994 commented 4 years ago

Hello, I want to reproduce your model, but I found that the quality is not as good as the model you provided. In your paper, you used batchsize = 16, but you used batchsize = 1 in your code. so which is the batchsize used by your model? I use the model you provided to predict my test set. The overall performance is very good and complete. the disp is as follow: image

However, if I use the model I trained to predict, I will lose some details (such as fingers). the disp is as follow: image

What do you think may caused this difference?

ShivamDuggal4 commented 4 years ago

HI @xubin1994 Glad that you liked our work. The batch size we used is 16, as we mentioned in the paper. Your result looks very good too. TBH, it's very hard to tell what could have caused such a minute difference. The results can always change slightly with re-training, with the GPU you used for training, with your environment (dependencies versions) etc.

Thanks, feel free to reopen if there is any other doubt.