Bootstrapping Mean Teacher with Proposal-Based Positive Sampling for Domain-Adaptive Object Detection

Object detection models often suffer performance degradation when applied to target domains with adverse weather conditions such as fog and rain due to domain shift. Domain-adaptive object detection (DAOD) addresses this issue without requiring annotations in the target domain, and the mean teacher (MT) framework is widely adopted as a baseline. However, the MT approach relies on pseudo-labels, which cause noise and instability, limiting its effectiveness under severe domain shift. In this work, we propose bootstrapping mean teacher (BMT), which is a non-contrastive learning framework designed to enhance target domain adaptation within the MT framework. BMT generates multiple positive samples from object proposals in both student and teacher networks and aligns them through a regression loss. To prevent representation collapse, BMT introduces architectural asymmetry by incorporating a non-linear multilayer perceptron into the student network, while the teacher is updated using an exponential moving average. BMT can be integrated seamlessly into existing DAOD pipelines without additional changes. Experimental results demonstrate its effectiveness under adverse weather scenarios. On the Cityscapes dataset to the Foggy Cityscapes dataset, BMT improves the CMT and AT by 0.3% and 1.3% mAP, respectively. On the CC-dataset, adapting from sunny to rainy conditions, it further enhances AT by 2.9% mAP, surpassing a detector fully supervised on the target domain.