search

JiaRenChang / PSMNet

Pyramid Stereo Matching Network (CVPR2018)
MIT License
1.43k stars 423 forks source link

regarding image normalization #10

Open stalin18 opened 6 years ago

stalin18 commented 6 years ago

First of all, congrats for this excellent work! Coming to my question, I think I never saw other papers explicitly mentioning image normalization for disparity estimation, do you have any numbers about how much image normalization improved your results? Secondly, why not use mean / stddev values from Scene Flow dataset instead of ImageNet? Thanks!

JiaRenChang commented 6 years ago

@stalin18 Hi, we tried normalizing images to [-1, 1] and using mean/std from Imagnet. The validation results are almost the same. The main reason why we use mean/std from ImageNet is that we want to finetune the model to natural images. The mean/std from a large natural image dataset may be better than from synthesis images.