[D] High-temperature softmax

I implemented a label propagation algorithm which is mainly used in the field of Video Object Segmentation (VOS). Basically I provide the labels for one frame and ask my model (using pre-trained encodings of frames) to do semantic segmentation on all the other frames of a video.

I am obtaining consistently better results using an high temperature softmax when computing the similarity between pixels of different frames. Then the top-k similarities of each pixel (features) are used to propagate the …

algorithm labels machinelearning propagation segmentation semantic softmax video