Paper on S2D: Sparse-To-Dense Keymask Distillation for Unsupervised Video Instance Segmentation accepted at NeurIPS 2026