VisTR

Why we included this project

VisTR is the official code for a CVPR 2021 paper that handles video instance segmentation in a single pass. Instead of detecting objects frame by frame and then linking the results across time, it runs a whole clip through a Transformer and gets back a set of instance sequences directly. That removes a lot of hand-built post-processing, and the design will feel familiar to anyone who has worked with DETR-style object queries. The repo ships training and inference scripts, pretrained weights, and YouTubeVIS setup instructions, so reproducing the reported numbers is straightforward. One caveat: training needs a big GPU (the authors used 32 GB V100s), so this is for teams with real compute, not quick experiments on a laptop.