Transfer Beyond the Field of View: Dense Panoramic Semantic Segmentation via Unsupervised Domain Adaptation

Autonomous vehicles clearly benefit from the expanded Field of View (FoV) 360° sensors, but modern semantic segmentation approaches rely heavily on annotated training data which is rarely available for panoramic images. We look at this problem perspective domain adaptation and bring to a setting, where labelled originates different distribution conventional xmlns:xlink="http://www.w3.org/1999/xlink">pinhole camera To achieve this, we formalize task unsupervised panoramic collect DensePass - novel densely dataset under cross-domain conditions, specifically built study Pinhole $\rightarrow$ PANORAMIC shift accompanied with pinhole examples obtained Cityscapes. covers both, labelled- unlabelled images, comprising 19 classes explicitly fit categories in source ( xmlns:xlink="http://www.w3.org/1999/xlink">i.e. pinhole) domain. Since data-driven models are especially susceptible changes distribution, introduce P2PDA generic framework Panoramic addresses challenge divergence variants attention-augmented modules, enabling transfer output-, feature-, feature confidence spaces. intertwines uncertainty-aware using values regulated on-the-fly through attention heads discrepant predictions. Our facilitates context exchange when learning correspondences dramatically improves performance accuracy- efficiency-focused models. Comprehensive experiments verify that our surpasses adaptation- specialized as well state-of-the-art methods.