About image normalization about densefusion

[greatwallet]

Hi, yes there's a bug but seems like the network can handle it quite well.

Hi!

So have you tried the standard way? For instance:

1. Scale the RGB image to [0., 1.] and then normalize using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225]?
2. Scale the RGB image to [0, 1] w/o ImageNet's normalization
3. Do neither, so that the RGB would be in range [0, 255].

The current dataloader gives out RGB in a really weird range, which bothers me a lot.

The above is visualization of the processed tensor from Linemod_preprocessed/data/01/rgb/0001.png using your dataloader pipeline. And its value range is [-2.1179, 1029.3066] w/o color-jittering. And this weirdly-ranged tensor is directly fed into network, which seems to be making little sense.

So should DenseFusion be working any better if I fix this to the standard way?

Thanks!

[j96w]

Hi, yes there's a bug but seems like the network can handle it quite well.