Transformers for Object Detection: A Comparative Study on a Large-Scale Autonomous Driving Dataset

Abstract

Recently, the transformer networks have emerged as a powerful tool for 2-D object detection from RGB camera images. This paper provides a comparative study on the performance of representative, state-of-the-art transformer-based object detection methods, including the Detection Transformer (DETR) and its variants, and demonstrates their performance and benchmark results on a new large-scale, real-world autonomous driving dataset. This paper also provides ablation studies on the effects of pre-training, post-processing, and cross-validation schemes on 2D object detection performance of transformer networks, which can be used to guide the selection and construction of transformer-based object detection algorithms based on their detection effectiveness and efficiency.