Open zhou-frank opened 4 years ago
@zhou-frank @feihuzhang Hello, I also found this problem when evaluating by code "evaluate.py" with the provided pre-trained model "sceneflow_epoch_10.pth". So, how to fix this problem in the evaluation?
Not sure if it's that same problem that I have #92, but looks similar. Any help?
I have found the reason of nan error,and willing to communicate with you via email. My email is xianda_guo@163.com
When I tested performance on sceneflow dataset, I encountered nan error in TEST/A/0149/left/0015.png. I found that this image's gt disparity pfm have problem that all of its disparity is large than max_disp(192).So a division by zero error will occur.
disp, height, width = readPFM("0015.pfm") print(disp)
The result of above code is zero.