NetGAN is a graph generative model; takes a graph and generates new graphs that resemble the input graph. By investigating components, the only necessary component is a low-rank approximation and all other components (incl. GAN) are not necessarily. proceedings.mlr.press/v119/rendsburg…
TransTrack is a transformer-based multiple-object tracking. Extract keys from a current frame, match them against queries from a previous frame for tracking, and match them against learned queries for new-coming object detection. Simple and effective arxiv.org/abs/2012.15460
They treat image segmentation as minimizing the mutual information between foreground and background and optimize the mask so that the foreground cannot be inpainted by background and vice versa. No learning required. SOTA on unsupervised segmentation. arxiv.org/abs/2012.07287