jzbontar / mc-cnn

Stereo Matching by Training a Convolutional Neural Network to Compare Image Patches
BSD 2-Clause "Simplified" License
707 stars 232 forks source link

Can not get same error rate which is referred on Kitti dataset #59

Closed MMMHT closed 5 years ago

MMMHT commented 5 years ago

Hi! mc-cnn is a amazing job! But I got some issue about the output of the model, neither the model trained by myself nor the model author provided can not produce error rate below 5%, in fact, I got both 11% error rate.

the command I used is : th main.lua kitti fast -a test_te -net_fname net/net_kittifast-a_train_all.t7

I didn't modify any part of code, just clone and run.

why would the accurate so different? what should I do to improve the accurate?

please help me!

MMMHT commented 5 years ago

32 works for me, it is the Nan problem caused by 0/0, modified to fill(1.0/1.0) will fix this problem