I am performing semantic segmentation. For that I have a dataset of videos in which every frame is labeled. Unfortunately, I don't know the order of the frames. That is why I am performing classic semantic segmentation using DeepLabV3+.
When inferring, the images are coming in a sequence (classic video). I thought maybe I can incorporate this information, by using predictions of earlier frames.
I know about semi-supervised video object segmentation, but I think this is a different use case. I just don't know to which category my problem belongs.
Methods such as MaskTrack incorporate knowledge of earlier frames, but they work on single images.
Does someone know a method, which uses a mask of an earlier frame to refine the mask of a consecutive frame?