Pyramid Transformer for Traffic Sign Detection

Automatic detection and classification of traffic signs have become an essential asset in the visual system of autonomous vehicles and self-driving cars. Recently, vision transformers have achieved remarkable performance on various benchmarks of visual tasks. We observe that prior ViTs could not provide satisfactory gain in traffic sign detection because the dataset size of this task is very small, and the class distribution of traffic signs is highly unbalanced. To solve the problems, we propose a novel Pyramid Transformer with hybrid architecture in this paper. Specifically, Pyramid Transformer follows a hierarchical architecture to build a feature pyramid with local and global information by using atrous convolutions. Furthermore, it inherits an intrinsic inductive bias and aims to learn multi-scale feature representation for objects of varying sizes, thereby enhancing the network robustness against the size discrepancy of traffic signs. We conduct experiments on the German Traffic Sign Detection Benchmark (GTSDB), which results demonstrate the superiority of the proposed model in the traffic sign detection task. More specifically, Pyramid Transformer achieves 77.8% mAP on GTSDB when applied to the Cascade RCNN as the backbone, which surpasses all the state-of-the-art methods.