HG-DETR: Image-Level Few-Shot Object Detection with Cross-Category and Query-Level Heterogeneous Graphs

Few-shot object detection FSOD aims to detect novel objects with limited annotated examples, yet existing methods face critical challenges in handling low-quality region proposals, leading to suboptimal generalization. And current meta-learning approaches often rely on pairwise region-class matching, which neglects contextual relationships among proposals and fails to leverage cross-class semantic dependencies, resulting in misclassification over similar classes and limited adaptability to novel categories. To address these limitations, we propose HG-DETR, a novel FSOD framework that integrates image-level detection with heterogeneous relational reasoning. Our method bypasses error-prone region proposal networks by directly operating on holistic image features through a Transformer-based architecture, enabling end-to-end optimization. By considering these multi-faceted relationships between proposals and classes, we propose 1 a cross-category semantic relationship graph that dynamically models semantic dependencies among base and novel classes to enhance prototype representations through knowledge transfer, 2 a query-level context aggregation graph models spatial relationships within a query image by connecting top-confidence proposals and a class node, using a GCN layer to aggregate features and refine proposals, and 3 bidirectional class-query adaptation via attention mechanisms to align feature distributions and bridge domain gaps. Qualitative and quantitative results demonstrate that our method achieves superior performance in few-shot object detection on Pascal VOC and MS COCO datasets compared with existing methods.