SlotMatch: Distilling Object-Centric Representations for Unsupervised Video Segmentation
SlotMatch is a simple knowledge distillation framework that transfers object-centric representations from a large teacher model to a lightweight student model for unsupervised video segmentation. It uses cosine similarity to align corresponding object slots between the two models, achieving superior performance with 3.6x fewer parameters and 2.7x faster inference than the teacher.