JiaRenChang
my val error is always equal to 0
hello, @JiaRenChang
I used the point cloud and the image to generate ground truth on the kitti raw data(the raw data about 600G), the result like this:
then I trained the model, but my 3-px error in val is always equal to 0.
why? can you help me? thank you very much!
@hnsywangxin Or you can plot the output disparity maps and compare with ground truth?
@JiaRenChang I'm so sorry to reply you now, my output disparity maps' pixels are all black, but I used kitti datasets to train,the output disparity maps' pixels is very good. so I suspected that there is a problem with the ground truth map I generated. But when I used another psmnet version(https://github.com/KinglittleQ/PSMNet) to train ,I can get a good disparity map result although not as good as yours. So why? I can't understand.
@JiaRenChang hello,I think it may be caused due to python version, because in https://github.com/KinglittleQ/PSMNet , python2.7 will cause the val error value to 0, but python3 is not
@hnsywangxin I guess that it was caused by some python packages. Or you can used python3 to train the model.
@JiaRenChang
I have used python3 to train the model,now,I use the model which was trained by myself can get the result that the pixels are not black,but, the val error value will also always equal to 0 .
I found in the code in the function of test in finetune.py
output3= model(imgL, imgR)
This value of output3 is a bit strange，my data get the value was very small,but for kitti data ,the value was very big
@hnsywangxin Hi, how do you fix the problem that error = 0.000? I also have this problem even in Python-3.8.
