Reduced-Resolution Head for Object Detection

Utilizing a multi-scale strategy to address the challenge of significant variations in object sizes is a common approach in object detection. However, this strategy often results in significant computational overhead on high-resolution feature maps, leading to increased model complexity. This paper proposes a novel Reduced-resolution Head for Object Detection (RHOD). It can be seamlessly incorporated into most existing popular detectors, operating in a plug-and-play way. By replacing high-resolution feature maps with lower-resolution ones for regression and classification, our RHOD effectively eliminates the computational costs on high-resolution feature maps. Combined with the proposed Feature Fusion (FF), the final detection accuracy of the model can approach the accuracy of the original detector. Experimental results demonstrate that RHOD reduces the computational cost of the detector by approximately 15%–40% without sacrificing accuracy. This reduction in computational cost leads to a significant improvement in inference speed, with an increase of 1.2–5.0 frames per second (FPS). The code and models will be available at  https://github.com/alpc128/RHOD .