Efficient Masked Attention Transformer for Few-Shot Classification and Segmentation
This work addresses few-shot classification and segmentation for computer vision applications, with incremental improvements focused on small objects and parameter efficiency.
The paper tackles the problem of few-shot classification and segmentation, where existing methods struggle with small objects, by proposing the Efficient Masked Attention Transformer (EMAT). EMAT improves accuracy for small objects and outperforms all methods on PASCAL-5^i and COCO-20^i datasets while using at least four times fewer trainable parameters.
Few-shot classification and segmentation (FS-CS) focuses on jointly performing multi-label classification and multi-class segmentation using few annotated examples. Although the current state of the art (SOTA) achieves high accuracy in both tasks, it struggles with small objects. To overcome this, we propose the Efficient Masked Attention Transformer (EMAT), which improves classification and segmentation accuracy, especially for small objects. EMAT introduces three modifications: a novel memory-efficient masked attention mechanism, a learnable downscaling strategy, and parameter-efficiency enhancements. EMAT outperforms all FS-CS methods on the PASCAL-5$^i$ and COCO-20$^i$ datasets, using at least four times fewer trainable parameters. Moreover, as the current FS-CS evaluation setting discards available annotations, despite their costly collection, we introduce two novel evaluation settings that consider these annotations to better reflect practical scenarios.