Enhancing Visual Domain Adaptation with Source Preparation
Published in arXiv, 2023
We address robotic perception in challenging environments like low-light scenarios by proposing Source Preparation (SP), a technique to reduce source domain biases. Our framework, Almost Unsupervised Domain Adaptation (AUDA), integrates three components: source preparation, unsupervised domain adaptation, and supervised alignment. We introduce CityIntensified, a new dataset with aligned image pairs from high-sensitivity and intensifier cameras for low-light semantic segmentation and object detection. Results demonstrate that SP improves performance across visual domains, achieving gains up to 40.64% in mIoU over baseline, while AUDA proves effective as a label-efficient approach requiring only limited target domain samples.