Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging

Natural scene semantic segmentation is an important task in computer vision. While training accurate models for semantic segmentation relies heavily on detailed and accurate pixel-level annotations, which are hard and time-consuming to be collected especially for complicated natural scenes. Weakly-supervised methods can reduce labeling cost greatly at the expense of significant performance degradation. In this paper, we explore the possibility of introducing hyperspectral imaging to improve the performance of weakly-supervised semantic segmentation. We take two challenging hyperspectral datasets of outdoor natural scenes as example, and randomly label dozens of points with semantic categories to conduct a point-supervised semantic segmentation benchmark. Then, a spectral and spatial fusion method is proposed to generate detailed pixel-level annotations, which are used to supervise the semantic segmentation models. With multiple experiments we find that hyperspectral information can be greatly helpful to point-supervised semantic segmentation as it is more distinctive than RGB. As a result, our proposed method with only point-supervision can achieve approximate performance of the fully-supervised method in many cases.