Disoccluded

Disocclusion is a term used in computer vision and graphics to describe the appearance of parts of a scene that become visible when the viewpoint changes, or the process of dealing with such regions in image synthesis and depth estimation. In stereo vision, disoccluded regions are image areas for which there is no corresponding pixel in the other view because a foreground object occludes them from one perspective but not another. In view synthesis, disocclusion creates holes in the synthesized image that correspond to surfaces revealed by shifting viewpoints.

Disoccluded regions pose a fundamental challenge because they lack direct correspondences across views. They can lead to artifacts in depth maps, 3D reconstructions, or rendered views if not properly handled. Approaches to manage disocclusion include filling holes with texture from surrounding areas, using temporal or multi-view information to infer plausible content, and applying priors or learned models to predict appearance and depth in the disoccluded zones.

Typical techniques range from traditional inpainting and patch-based texture synthesis to geometry-aware methods that propagate depth and color from neighboring pixels, as well as modern deep learning approaches that generate plausible content for disoccluded regions. The choice of method often depends on the application, such as real-time rendering, 3D reconstruction, or video coding, and on the available input data like multiple views or temporal frames.