Low Rank Adaptation for Stable Domain Adaptation of Vision Transformers

Unsupervised domain adaptation plays a crucial role in semantic segmentation tasks due to the high cost of annotating data. Existing approaches often rely on large transformer models and momentum networks to stabilize and improve the self-training process. In this study, we investigate the applicability of low-rank adaptation (LoRA) to domain adaptation in computer vision. Our focus is on the unsupervised domain adaptation task of semantic segmentation, which requires adapting models from a synthetic dataset (GTA5) to a real-world dataset (City-scapes). We employ the Swin Transformer as the feature extractor and TransDA domain adaptation framework. Through experiments, we demonstrate that LoRA effectively stabilizes the self-training process, achieving similar training dynamics to the exponentially moving average (EMA) mechanism. Moreover, LoRA provides comparable metrics to EMA under the same limited computation budget. In GTA5 → Cityscapes experiments, the adaptation pipeline with LoRA achieves a mIoU of 0.515, slightly surpassing the EMA baseline’s mIoU of 0.513, while also offering an 11% speedup in training time and video memory saving. These re-sults highlight LoRA as a promising approach for domain adaptation in computer vision, offering a viable alternative to momentum networks which also saves computational resources.