search

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

What are the correct dimension and ndisp values for Middlebury datasets? #56

Open tltsilveira opened 6 years ago

tltsilveira commented 6 years ago

Hey everybody. I am trying to predict depth from the test images of Middlebury 2006. However the obtained results are very poor and they have a large portion of estimated map the with no disparity (like the maximum value).

In particular I am using the mb slow model on full size images.

ndisp is set as ndisp = max(gtDisp) * 1.1 where gtDisp is the ground-truth disparity map for of a given stereo pair of interest.

Do you know this issue?

Thank you in advance.

saeid-h commented 6 years ago

if you open the calib.txt file for each image, you will find it. Files contain information like this: cam0=[1038.018 0 322.037; 0 1038.018 243.393; 0 0 1] cam1=[1038.018 0 375.308; 0 1038.018 243.393; 0 0 1] doffs=53.271 baseline=176.252 width=718 height=496 ndisp=73 isint=0 vmin=8 vmax=65 dyavg=0.184 dymax=0.423