Problem about the results compared with Middlebury displayed results

[shuluoshu]

Hi,
Thanks for your code, and after I ran the slow network to predict the trainingH datasets, I got much worser results than the results displayed on the Middlebury website, I wonder where I did something wrong, can you help me out ?
The command is as follows:
./main.lua mb slow -a predict -net_fname net/net_mb_slow_-a_train_all.t7 -left im0.png -right im1.png -disp_max 190
And one of the results can be compared as follows: The error rate(nocc) for Vintage by running the code is 34% while the result of Vintage image on Middlebury (MC-CNN-acrt ) is 24.8%.
Do I miss something important? Thanks a lot !

[jzbontar]

I think you didn't set the -disp_max parameter correctly. I believe it should be 380 (=760 / 2) for the Vintage stereo pair.

[shuluoshu]

Thanks a lot, sorry for that~

[wujunjie15]

@shuluoshu Hello, i set disp_max to 380. Unlike the official result, the error is still very large. Can you solve this problem? if you address the problem, please tell me some detail.