3D reconstruction of indoor scenes with implicit representation

We propose a novel framework for reconstructing comprehensive three-dimensional indoor scenes, which includes both room backgrounds and indoor objects, based on single-view images. Our method predicts object shapes, poses, and scene layouts. Existing approaches often struggle with accurately reconstructing 3D indoor scenes due to the challenges posed by heavy occlusion. While they can generate 3D shapes of indoor objects, their emphasis on geometric shape recovery frequently overlooks critical aspects of object appearance and fine details. This limitation leads to inaccuracies in estimating shapes and layouts, especially in cluttered indoor environments where objects are often heavily occluded. To tackle this issue, we introduce depth implicit representation as a means to enhance the accuracy and fidelity of indoor scene reconstruction.