Active Domain Adaptation with Multi-level Contrastive Units for Semantic Segmentation

To further reduce the cost of semi-supervised domain adaptation (SSDA) labeling, a more effective way is to use active learning (AL) annotate selected subset with specific properties. However, tasks are always addressed in two interactive aspects: transfer and enhancement discrimination, which requires data be both uncertain under model diverse feature space. Contrary classification tasks, it usually challenging select pixels that contain above properties segmentation leading complex design pixel selection strategy. address such an issue, we propose novel Active Domain Adaptation scheme Multi-level Contrastive Units (ADA-MCU) for semantic image segmentation. A simple strategy followed construction multi-level contrastive units introduced optimize supervised learning. In practice, MCUs constructed from intra-image, cross-image, cross-domain levels by using labeled unlabeled pixels. At each level, define losses center-to-center pixel-to-pixel manners, aim jointly aligning category centers reducing outliers near decision boundaries. addition, also introduce categories correlation matrix implicitly describe relationship between categories, used adjust weights MCUs. Extensive experimental results on standard benchmarks show proposed method achieves competitive performance against state-of-the-art SSDA methods 50% fewer significantly outperforms large margin same level annotation cost. Code will https://github.com/haoz19/ADA-MCU .