Training with random crop

[Archerwz]

Hi, I'm trying to combine monodepth with other stereo methods, since most of stereo depth prediction is using random crop when training, so I use random crop on monodepth together with stereo methods. However, the mono part in my network is not working, the whole network is just the stereo part working solely. So I was wondering if adding supervision with mono loss will deal with this, just like your works here.
My case is training with random crop with size 256*512, when testing, I feed the full resolution image. What's your opinion about using random crop when training semiDepth?

[jahaniam]

Hi,

I'm almost sure you can do the random cropping. Just make sure you crop both sides (left and right) equally.
Random cropping shouldn't really matter much with either monodepth or other stereo methods except training might converge faster due to direct sparse supervision. I think both monodepth and this method should work with random crops.

[Archerwz]

Thank you, I'll check whether there's a bug in my code. And also, since the output of monodepth is normalized along the width, so when adding supervision, you have to scale the disparity map with image width. For resize it's simple just multiply by lets say 1242, but when using random crop, during training phase, the input image is 256x512, so before supervision, you have to multiply by 512, but when testing, the input image is in its full resolution 375x1242, do I multiply by 1242 or 512?