D2D: Divide to Detect, A Scale-Aware Framework for On-Road Object Detection Using IR Camera

In this paper, to solve the problem of inconsistencies between the predictions in today's SOTA object detection networks, which incorporate the pyramid architecture with multi-level prediction, we proposed a scale-aware framework for IR image-based on-road object detection. The proposed framework uses scale-based attention mechanism to assign responsibilities to each feature levels. With this design, each feature level will focus on detecting a certain range of object scales, thereby minimizing the conflict among the predictions in the final result. Compared to Scaled-YOLOv4 baseline, our proposed method can achieve better performance without increasing FPS on FLIR dataset. The experimental results on RGB image-based object detection datasets also show that our proposed method gives good improvements when applied to RGB images.