Rethinking Cross-Domain Pedestrian Detection: A Background-Focused Distribution Alignment Framework for One-Stage Detectors

Yancheng Cai, Bo Zhang, Baopu Li, Tao Chen, Hongliang Yan, Jingdong Zhang

Abstract: Cross-domain pedestrian detection aims to generalize pedestrian detectors from one label-rich environment to another label-scarce environment, which is vital in enormous realworld applications. Recent works generally rely on domain alignment to train domain-adaptive detectors, either on image-level or instance-level. Due to the proposal-free rapid detection, we focus on the one-stage domain-adaptive detector design in this work. We find that the lack of instance-level proposals for the one-stage detector makes it only be able to do image-level feature alignment, causing the foreground-background misalignment issue, that is, the foreground features in the source domain image are falsely aligned with background features in the target domain image. To resolve the conflict between foreground and background in the alignment stage, we systematically analyze the importance of foreground and background in image-level cross-domain alignment, and learn that background plays a more important role in imagelevel cross-domain alignment. Therefore, we focus on background cross-domain feature alignment while minimizing the influence of foreground features on the cross-domain alignment stage. This paper proposes a novel Background-focused Distribution Alignment Framework (BFDA) to train domain adaptive onestage pedestrian detectors. Specifically, BFDA first decouples the background features from the whole image feature maps and then aligns them via a novel long-short-range discriminator. Extensive experiments show that BFDA significantly enhances the crossdomain pedestrian detection performance, compared with the mainstream domain adaptation technologies for either one-stage or two-stage detectors. Moreover, by employing the efficient onestage detector (YOLOv5) as the backbone, BFDA can reach 217.4 FPS (640×480 pixels) on NVIDIA Tesla V100 (7∼12 times FPS of the existing frameworks), which is very meaningful for practical applications. The code will be made publicly available.