COQ-DETR: Category-Aware Object Queries in Meta-DETR for Few-Shot Object Detection

Few-shot object detection (FSOD) aims to detect objects in an image with only a few annotated samples. Among existing methods, Meta-DETR has proved to be particularly effective for FSOD. Meta-DETR adapts the DEtection TRansformer (DETR) framework by leveraging support classes to generate category-specific support-aggregated query features exploited by a set of learnable object queries for final detection. However, Meta-DETR exhibits a key weakness: while it effectively generates support-aggregated features, its object queries may not adapt well to novel classes, resulting in suboptimal detection performance. To address this issue, we propose a novel approach that embeds support class information directly into the object queries, thus enhancing the adaptability of the detection model to novel classes, resulting in improved detection accuracy. Experimental results demonstrate that our method outperforms Meta-DETR by up to 26.74% for Pascal VOC and 16.5% for MS COCO, offering a more robust solution for few-shot object detection.