Object Detection for Traffic Scenarios Based on Scale-Aware Label Assignment and Dynamic Class Suppression Loss

For object detection in traffic scenarios, small-sized and long-tail distributed objects are major challenges. However, most previous studies consider these two important problems as irrelevant issues. In addition, these works achieve improvements at the cost of significant computation increase. In this article, we unveil that both object size and frequency should be taken into consideration in a unified manner. We first propose a scale-aware label assignment strategy to focus on relatively small objects; and it intuitively adjusts sampling areas according to the target size. Based on the assignment results, a dynamic class suppression loss is designed to boost the performance on tail categories from a statistic-free perspective. We conduct extensive experiments on typical benchmarks; results show that the proposed method achieves more than 6.5% improvements over the baseline. Without aggravating the computation burden, our method effectively enhances object detection performance in traffic scenarios.