OnlineDataset Issues about cascadepsp

[hkchengrex]

It is intended. Note that one of our key features is that we only need to train on low-resolution images with the largest being around 400x500 while extending to high-resolution only in evaluation. You should not be using the online dataset for large images. If you have sufficiently many annotated high-resolution training images, this work is probably not the most suitable (GLNet might be better in that situation).

[hkchengrex]

Thanks for reporting.

For 2 it is probably due to the new torchvision version, see pytorch/vision#2520.
One quick fix is to revert to torchvision 0.6.0.

I will probably fix these soon.

[urimerhav]

Got it.

By the way, it looks like the way the code handles the dataset, you look at random crops of 224 pixels. This is surprising because the meaning of 224 pixels may be either very broad context if the original image is pretty small, or a tiny patch takes completely out of context if the original image was, say, 4000X4000 pixels. Is this intended behavior?

[urimerhav]

Right. It's certainly counter-intuitive that training on one scale and predicting on another would yield such impressive results! I guess I need to reread what you wrote on this in more depth.

[hkchengrex]

Fixed.