Hyperbolic Active Learning for Semantic Segmentation under Domain Shift

We introduce a hyperbolic neural network approach to pixel-level active learning for semantic segmentation, and propose a novel geometric interpretation of the hyperbolic geometry that arises bottom-up from the statistics of the data. In our formulation the hyperbolic radius emerges as an estimator of the unexplained class complexity, which encompasses the class intrinsic complexity and its scarcity in the dataset. The unexplained class complexity serves as a metric indicating the likelihood that acquiring a particular pixel would contribute to enhancing the data information. We combine this quantity with prediction uncertainty to compute an acquisition score that identifies the most informative pixels for oracle annotation. Our proposed HALO (Hyperbolic Active Learning Optimization) sets a new state-of-the-art in active learning for semantic segmentation under domain shift, and surpasses the supervised domain adaptation performance while only using a small portion of labels (i.e., 1%). We perform extensive experimental analysis based on two established benchmarks, i.e. GTAV $\rightarrow$ Cityscapes and SYNTHIA $\rightarrow$ Cityscapes, and we additionally test on Cityscape $\rightarrow$ ACDC under adverse weather conditions.