Visiting the Invisible: Layer-by-Layer Completed Scene Decomposition

Existing scene understanding systems mainly focus on recognizing the visible parts of a scene, ignoring intact appearance physical objects in real-world. Concurrently, image completion has aimed to create plausible for invisible regions, but requires manual mask as input. In this work, we propose higher-level system tackle both and backgrounds given scene. Particularly, built decompose into individual objects, infer their underlying occlusion relationships, even automatically learn which are occluded that need be completed. order disentangle relationships all complex use fact front object without being is easy identified, detected, segmented. Our interleaves two tasks instance segmentation through multiple iterations, solving layer-by-layer. We first provide thorough experiment using new realistically rendered dataset with ground-truths regions. To bridge domain gap real imagery where unavailable, then train another model pseudo-ground-truths generated from our trained synthesis model. demonstrate results wide variety datasets show significant improvement over state-of-the-art.