Efficient Masked Attention Transformer for Few-Shot Classification and Segmentation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Carrión-Ojeda, Dustin, Roth, Stefan, Schaub-Meyer, Simone
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911416953339904
author Carrión-Ojeda, Dustin
Roth, Stefan
Schaub-Meyer, Simone
author_facet Carrión-Ojeda, Dustin
Roth, Stefan
Schaub-Meyer, Simone
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2507_23642
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Efficient Masked Attention Transformer for Few-Shot Classification and Segmentation
Carrión-Ojeda, Dustin
Roth, Stefan
Schaub-Meyer, Simone
Computer Vision and Pattern Recognition
Artificial Intelligence
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.
title Efficient Masked Attention Transformer for Few-Shot Classification and Segmentation
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2507.23642