PSMNet
PSMNet copied to clipboard
JiaRenChang
🔥🔥🔥some questions?
1、why sceneflow dataset need to pad in the test, but kitti2015 don't need?
2、code about computing 3-px error
true_disp = disp_true
their memory address are same, so
correct = (disp_true[index[0][:], index[1][:], index[2][:]] < 3) | ( disp_true[index[0][:], index[1][:], index[2][:]] < true_disp[ index[0][:], index[1][:], index[2][:]] * 0.05) is wrong.
3、in the finetune.py,
if total_test_loss / len(TestImgLoader) * 100 > max_acc: max_acc = total_test_loss / len(TestImgLoader) * 100 max_epo = epoch
you test return the error between the pred and GT，but when you in the main() you see the max error as max acc, why? is it the error bigger the model' weight is worse. May you explain the reason?
4、why in the main, you set the mask below the args.maxdisp; but in the finetune, you set the mask up the 0? why?
3、in the finetune.py,
if total_test_loss / len(TestImgLoader) * 100 > max_acc: max_acc = total_test_loss / len(TestImgLoader) * 100 max_epo = epoch
you test return the error between the pred and GT，but when you in the main() you see the max error as max acc, why? is it the error bigger the model' weight is worse. May you explain the reason?
I aslo notice this, and I am confused about it ! @JiaRenChang
