I have a question about "additional information".
Challenge rule says "For this track only the images are to be used. The use of any other additional information (semantics, object instances) is not allowed."
Does it means that any kinds of supervision except "images (RGB) themselves", are not allowed?
For example, perceptual loss (Johnson et al.) uses ImageNet pre-trained VGG features.
It can be treated as classification supervision.
Hello taeohkim,
You can use Perceptual Loss for training your network. You can also train with images from other datasets, but you are not to use the ADE20k's validation set, as a subset of it will be used for our own validation set.
Best,
Evan