Here we demonstrate the inverse rendering capabilities of PANDORA on real datasets. Please view this html preferably in Chrome browser. The slider can be moved by hovering mouse over the video. Any video can be paused by clicking on the video. The video can then be resumed by refreshing the page.

Comparison with PhySG on Captured Scenes

PANDORA reconstructs more detailed surface geometry compared to state-of-the-art inverse rendering technique PhySG, such as, on the nose and the bird. PANDORA also accurately models the broad specular highlights on this relatively rough surface, such as on the figurine's hat. PANDORA also outputs a plausible illumination estimation that captures the major light sources.

Surface Normals

Net Radiance Field

Diffuse Radiance Field

Specular Radiance Field

Estimated Illumination

Ground Truth Illumination

overview

Role of polarization cues

Here compare our performance with the same model trained only on intensity images and without any polarimetric cues. Polarimetric cues aide us in performing more accurate diffuse-specular separation and recovers cleaner geometry. The model without polarization has texture artefacts in the normals and the specular, such as on the owl's head and back respectively. Moreover, the model without polarization fails to capture small lights in the incident illumination.

Surface Normals

Net Radiance Field

Diffuse Radiance Field

Specular Radiance Field

Estimated Illumination

Ground Truth Illumination

overview

Appearance Editing

PANDORA's radiance decomposition can be used for editing the radiance field such as by changing the color of the diffuse component without affecting the specular highlights or for imparting a metallic appearance to the object.

Original Appearance

Changing diffuse color

Metallic Apperance