Capturing, Editing, and Relighting Indoor Scenes from a Single Panorama

Virtual staging has been widely used to showcase indoor spaces. However, architectural design requires the integration of disparate software tools and significant manual input, which limits its scalability across various spaces. Recent computer vision applications have started using a single photograph to understand 3D scenes. This thesis introduces a novel inverse rendering pipeline that automatically transforms an indoor panorama into a new virtual panorama within its outdoor context.

The first contribution is an indoor-outdoor HDR photography method. In each scenario, an indoor panorama is captured along with its corresponding outdoor image to understand indoor spatially varying light. This effort has resulted in a new HDR dataset that includes 278 scenes with photometric calibration.

The second contribution is a full inverse rendering pipeline. Conventional data-driven approaches face limitations in constructing accurate global illumination for indoor scenes. In this work, the 3D floor layout, material properties, and illumination conditions are directly estimated from the captured photographs, and a virtual model is constructed for realistic scene editing and relighting.

The third contribution is a light-to-heat estimation method. Beyond the pixel, the absolute light level for virtual indoor scenes is computed to guide indoor activities. Meanwhile, transient heat simulation is implemented to estimate the heat generated by natural light on indoor surfaces.