DiDA: Video Object Segmentation with Distillation Learning of Deformable Attention
This research introduces DiDA, a novel self-supervised video object segmentation method that combines deformable attention and knowledge distillation. The goal was to create a lightweight architecture that adapts effectively to temporal changes in video while maintaining high performance. The method achieves this by using deformable attention for flexible temporal adaptation and a new distillation framework that transfers both attention maps and logits, making VOS more efficient for low-power devices.