Image Segmentation Network, strange looking images

[ThomasDelteil]

Using gluoncv 0.2 this is the output I get when running the image segmentation tutorial:

Is this the expected output? The mask looks strange and the denormalized image has artefact.

[zhreshold]

@zhanghang1989

[ThomasDelteil]

The reason is because the training images are augmented by default. Discussed with @hetong007 about that and in my opinion, there should be a flag to disable the default augmentation, or default augmentation should be passed by default to the transform argument. Right now you would need to do

trainset = gluoncv.data.VOCSegmentation(split='train')
trainset.mode = 'test'

in order to get training images without data augmentation.

[ThomasDelteil]

the artefact happen also on validation set, root cause was a bug in denormalize https://github.com/dmlc/gluon-cv/pull/181

[zhanghang1989]

Thanks for reporting! There was a bug in denormalization.