ReDet: Effective Real-time Object Detection via Efficient Multi-scale Extraction Aggregation

Jian Li, Xin Jiang, Lu Jin, Zechao Li

Abstract: Real-time object detection demands detectors that excel in both speed and accuracy. However, existing methods rely on complex Feature Pyramid Networks and computationally intensive post-processing to boost performance, often struggling to balance efficiency and accuracy. In this paper, we propose ReDet, an efficient real-time end-to-end object detection framework that improves detection capability while maintaining low computational cost. Specifically, we propose a Multi-scale Extraction Aggregation module to enhance feature fusion ability with minimal overhead, boosting feature representation across scales. Additionally, a Regression Enhancement Module is incorporated to mitigate the assignment inconsistency between localization and classification, further enhancing detection accuracy with negligible computational cost. Moreover, we introduce a Multi-label Auxiliary Strategy to eliminate the reliance on post-processing by enabling one-to-one label assignment. The experimental results demonstrate that ReDet-T achieves 39.2% AP and 500 FPS on the COCO val2017 dataset, while ReDet-S achieves 45.3% AP and 294 FPS, outperforming many other detectors in both speed and accuracy.

External IDs:dblp:conf/icmcs/LiJJL25