KD-Former: Transformer Knowledge Distillation for Image Matting

Vision transformers (ViTs) have been outstanding in multiple dense prediction tasks, including image matting. However, the high computational and training costs of ViTs lead to a bottleneck for applications on low computing power devices. In this paper, we propose a novel transformer-specific knowledge distillation (KD-Former) framework for image matting that can effectively transfer core attribute information to improve the lightweight transformer model. To enhance the information transfer effectiveness in each stage of Vits, we rethink transformer knowledge distillation via dual attribute distillation modules - Token Embedding Alignment (TEA) and Cross-Level Feature Distillation (CLFD). Extensive experiments demonstrate the effectiveness of our KD-Former framework and each proposed key component. Our lightweight transformer-based model outperforms the state-of-the-art (SOTA) matting models on multiple datasets.