Image segmentation on the Oxford-iiit pet dataset using CNNs, GNNs, and U-Nets. The basic CNN is only there to test the code. The goal was to test the effectiveness of SCG-Net: Self-Constructing Graph Neural Networks for Semantic Segmentation. The version with the variational auto-encoder was very bad and basically did just predict one class. Using a normal auto-encoder improved the result but it was still not good enough for me, also the interpolation used led to a very non-smooth image as can be seen below(sorry for the mouse but I am too lazy to redo the img).
This brought me to try U-Net too. The results of this model were acceptable as can be seen below:
All methods could have worked better with longer training times, but my hardware and patience are limited. The weights of SCG-Net are in the repo, the U-Net weights take up too much space. I have also changed/improved U-Nets architecture successfully, but the model is for now not to be published because of a university project.