FII-DETR: Few-shot object detection with fully information interaction

Abstract

Few-shot object detection (FSOD) aims to effectively classify and localize objects in images with only a few annotated samples. Recent meta-learning-based DETR approaches achieve promising performance in FSOD tasks. However, classifying confusing categories remains a critical challenge, particularly in scenarios involving occluded or small objects. To tackle this problem, we propose a meta-learning FSOD model built upon Deformable DETR, focusing on full information interaction, named FII-DETR. Firstly, an Adaptive Foreground Enhancement (AFE) module is designed to adaptively enhance important information and edge-aware representations in support images, enabling the model to capture discriminative features more effectively. Secondly, a Multiscale Local Information Fusion (MLIF) module and a Global Symmetric Aggregation (GSA) module are proposed to enhance local information interaction and aggregate support and query features from local and global perspectives. In addition, we introduce self-supervised pretraining (SSP) into the meta-learning FSOD framework to further enhance FII-DETR’s generalization capability by maximizing the mutual information of prior knowledge. We comprehensively evaluate the performance of FII-DETR on PASCAL VOC and MS COCO benchmarks. FII-DETR outperforms state-of-the-art FM-FSOD by 3 %, Meta-DeDETR by 2.6 %, and Meta-DETR by 6.5 %, averaging three splits on the PASCAL VOC. On the COCO dataset, FII-DETR outperforms Meta-DETR and Meta-DeDETR and is also superior to FM-FSOD in the 1-shot and 3-shot settings. This work demonstrates that fully information interaction and aggregation can provide effective and robust support for improving the performance of FSOD built upon DETR.