DSF-DETR:Real-Time End-to-End Object Detection with Multi-scale Fusion and Lightweight Activation

Object detection has witnessed significant advancements, especially in real-time detectors such as the YOLO series and end-to-end transformer-based models. However, YOLO models rely on Non-Maximum Suppression (NMS) for post-processing, which hinders their deployment flexibility. Meanwhile, transformer detectors such as DETR suffer from high computational costs. This paper proposes DSF-DETR, a model built upon the RT-DETR framework. It incorporates a lightweight depth-separable fusion mechanism to enable RT-DETR to more effectively utilize multi-scale features from ResNet-50. Moreover, to accelerate model inference, we employ a lower-complexity Telu activation function. The experiment on the COCO 2017 dataset shows that the proposed method preserves high accuracy while exploring a more efficient multi-scale fusion mechanism, demonstrating a viable trade-off between architectural simplicity, efficiency, and performance.