Teacher–student adversarial YOLO for domain adaptive detection in traffic scenes under adverse weather

Abstract

The problem of difficult traffic object localization under adverse weather has not been solved due to the labor-intensive and time-consuming process of collecting and labeling large-scale data. Domain adaptive object detection (DAOD) can achieve cross-domain detection without labeling, however, most of the existing DAOD methods are based on two-stage Faster R-CNN, which needs to be improved in both accuracy and speed. We propose a DAOD method TSA-YOLO, which takes full advantage of adversarial learning and pseudo-labeling to achieve high-performance cross-domain detection for fog, rain, and low-light scenes. For the input images, we generate auxiliary domain images by CycleGAN, also design a strong and weak enhancement method to reduce the bias of the teacher and student models. Additionally, in the student self-learning module, we propose a pixel-level domain discriminator to better extract domain-invariant features, effectively narrowing the feature distribution gap between the source and target domains. In the teacher–student mutual learning module, we incorporate the mean teacher (MT) model, iteratively update parameters to generate high-quality pseudo-labels. In addition, we evaluate our method on the public datasets Foggy Cityscapes, Rain Cityscapes, and BDD100k_Dark. The results show that TSA-YOLO significantly improves detection performance. Specifically, compared with the baseline, TSA-YOLO achieves up to a 15.0% increase in [email protected] on Foggy Cityscapes and up to an 18.5% increase on Rain Cityscapes, while converging in only 50 epochs and without reducing the model’s inference speed.