QueenVIS · Video Instance Segmentation

Detect, segment and track every object instance across a video clip — from a model that never saw a single annotated video during training. QueenVIS enriches Mask2Former object queries with auxiliary feature/center supervision on still images, then associates instances at test time with training-free confidence-guided query propagation plus a non-parametric memory bank (zero extra parameters, zero extra GFLOPs).

Paper · Project page · Code · Weights

_{QueenVIS weights are released under the NVIDIA Source Code License-NC — non-commercial research and evaluation only. This Space runs the deformable-attention op in pure PyTorch (the custom CUDA kernel cannot be compiled here), which is numerically equivalent but slower.}