search

JiaRenChang / PSMNet

Pyramid Stereo Matching Network (CVPR2018)
MIT License
1.43k stars 423 forks source link

pretrained_sceneflow model error #169

Open Longseabear opened 4 years ago

Longseabear commented 4 years ago

Hello. I'm working on stereo matching. I found two problems in the provided code. the pretrained sceneflow model seems odd. It shows consistently different results from ground-truth. The mean error increases in proportion to the disparity scale. When I trained again, it didn't matter. I think you need to modify the pretrained weight.

jhk-chd commented 4 years ago

I also come across the same problem, when I download the pretrained scene flow model and test it on the scene flow dataset, the EPE error is large (larger than 6.0). It's totally different from the 1.09. I also think you need to check the pretrained weight of the pretrained scene flow model

Sarah20187 commented 4 years ago

@Longseabear @JunhuaKang I had the same problem at the very beginning. I think it might be caused by the version of some python packages.

jhk-chd commented 4 years ago

@Longseabear @JunhuaKang I had the same problem at the very beginning. I think it might be caused by the version of some python packages.

Hi, thanks. Unfortunately, I have tried using the different version of Numpy, it didn't work. I still got the EPE error of Scene Flow larger than 6 pixels. Can you show the environment you used?

Sarah20187 commented 4 years ago

@Longseabear @JunhuaKang I had the same problem at the very beginning. I think it might be caused by the version of some python packages.

Hi, thanks. Unfortunately, I have tried using the different version of Numpy, it didn't work. I still got the EPE error of Scene Flow larger than 6 pixels. Can you show the environment you used?

FYI. https://github.com/JiaRenChang/PSMNet/issues/64

jhk-chd commented 4 years ago

@Sarah20187 Hi, It didn't work either. Did you get the reported error by just testing their pre-trained model on Scene flow or did you train it by yourself?

XinyiYing commented 4 years ago

I come across the same problem. Is there any solutions?