disocclusionaware

Disocclusionaware refers to techniques and models that explicitly account for disocclusion in visual processing tasks. Disocclusion occurs when parts of a scene become newly visible as the viewpoint changes or as objects move, leaving regions without a direct correspondence to previous frames or views. Disocclusionaware methods aim to handle these regions correctly rather than treating them as mere missing data.

In computer vision, disocclusion awareness is particularly important for stereo matching, multi-view reconstruction, and view synthesis. In stereo, disoccluded areas lack reliable correspondences, which can lead to artifacts if not handled properly. In video and view synthesis, newly revealed regions must be filled plausibly to maintain realism and temporal coherence. Depth sensing and depth completion also benefit from disocclusion-aware modeling to produce consistent depth maps in the presence of occlusions.

Common approaches include learning to predict occlusion masks alongside primary estimates (such as depth or disparity), incorporating left-right consistency checks, and using inpainting or texture synthesis to fill disoccluded regions. Temporal priors, motion models, and explicit visibility reasoning are used to improve stability across frames. Some methods integrate occlusion reasoning into end-to-end networks, training with synthetic data or specially annotated datasets to supervise both depth and occlusion predictions.

Challenges remain in accurately detecting small or gradual occlusions, preserving sharp boundaries, and maintaining real-time performance. Disocclusionaware techniques continue to advance the realism and accuracy of 3D reconstruction and view synthesis.