OccluFree: Occlusion-Aware Restoration for Amodal Building Windows and Doors Segmentation

Abstract

Urban applications, such as facade reconstruction, digital twins, building analysis, and semantic 3-D city modeling benefit from reliable and geometrically consistent representations of window and door (WD) elements. Many image-based facade reconstruction pipelines first detect WD elements in street-level imagery before integrating them into downstream 3-D reconstruction. However, these pipelines often do not explicitly account for occlusions during image-level detection, which can lead to incomplete element boundaries and propagate errors into subsequent modeling stages. Since WD elements are frequently occluded by vegetation, vehicles, and street furniture, recovering complete facade geometry from a single image remains challenging. Although multiview reconstruction and 3-D completion methods can exploit geometric redundancy across viewpoints, persistent occlusions, or limited image coverage may still result in incomplete observations. To address this challenge, we propose OccluFree, an image-level framework that jointly localizes and restores occluded facade elements, producing amodal segmentation masks that provide cleaner observations for downstream 3-D reconstruction and modeling. OccluFree employs a dual-branch occlusion localization module to generate pixel-level occlusion masks across diverse facade architectures. These masks guide an occlusion-aware restoration module, which is fine-tuned on the UFE dataset, comprising approximately 137K full facade and WD element images that span varied occlusion scenarios and architectural styles. Validation on the multi-fold occlusion (MFO) benchmark of approximately 19K images yields 82% mIoU, 88% $\text{mF}1$, and 79% mean overall accuracy, demonstrating the potential of the proposed framework to improve facade completeness and provide higher quality inputs for downstream urban modeling applications.