search

JiaRenChang / PSMNet

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

Losses in training have a lot of shock #104

Open walkingmelon opened 5 years ago

walkingmelon commented 5 years ago

Dear@JiaRenChang ，When I used PSMNet to train SceneFlow datasets, the loss was quite volatile. Is the data set not clean? And how do I deal with it? Should I delete some unreliable data? Look forward to your reply！

JiaRenChang commented 5 years ago

@walkingmelon Some image pairs in Scene Flow have very large disparities. We have masked the disparities are larger than 192 (line 84 in main.py). You may verify the code.

walkingmelon commented 5 years ago

Thank you for your timely reply. I also noticed the restrictions on disparity in the code, and I didn't modify the code. Attached picture is my training loss（partial）. 1773168584

JiaRenChang commented 5 years ago

@walkingmelon Yep, it is fluctuating. Which pass of Scene Flow do you use ? We use the cleanpass version.

walkingmelon commented 5 years ago

Yes, I use the same data set. Maybe my data set is contaminated. I will now download the dataset again. I hope I will have good results.