max-manning/autoregunsupseg

Reimplementation of "Autoregressive Unsupervised Image Segmentation" by Ouali et al (ECCV 2020)

autoregunsupseg's Issues

How to apply this to new image or new Datasets

How can I un-resize the image after get the segmentation result?

In train and validate, the image is loaded in 3128128 size, so the segmentation result has batchsize1128*128 shape. How can I un-resize the result, so that I can get the seg result images just like the images shown in the paper？

Looking forward to your reply, thanks very much!

How can the model differentiate Ground、Sky、Plants、Non-stuff after training？

I'm very curious about that how the model have the ablity to differentiate Ground、Sky、Plants、Non-stuff without explicit labels but with just no-label training? How does the model know to partition these kinds of parts rather than such as human-animal-plant partition? How does the model do the segment-label map(This question, in my mind, maybe implemented by the Hungarian algorithm)?

Looking forward to your reply!!