search

dontLoveBugs / DORN_pytorch

PyTorch implementation of Deep Ordinal Regression Network for Monocular Depth Estimation
298 stars 67 forks source link

question about transform gt (depth map) #13

Closed cai199626 closed 4 years ago

cai199626 commented 5 years ago

In kitti dataloader

s = np.random.uniform(1.0, 1.5)  # random scaling
transform = my_transforms.Compose([
            my_transforms.Crop(130, 10, 240, 1200),
            my_transforms.Resize(460 / 240, interpolation='bilinear'),
            my_transforms.Rotate(angle),
            my_transforms.Resize(s),
            my_transforms.CenterCrop(self.size),
            my_transforms.HorizontalFlip(do_flip)
        ])
gt_ = transform(gt)
gt_ /= 100.0 * s

I wonder why gt should divide by s(the random scale factor). If so, there are two resize opts in transform, but only divide s. The first scale factor 460/240 is ignored, why? The same problem in nyu dataloder. Thanks a lot!

dontLoveBugs commented 5 years ago

Yes, if I have time, I will correct it.