Photometric Ambient Occlusion for Intrinsic Image Decomposition.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Ambient Occlusion (AO) quantifies local visibility in scenes, crucial for realistic rendering but underexplored in computer vision.
• Existing methods for AO computation often require complex geometric data or lack applicability to real-world image stacks.

Purpose of the Study:

• To develop a novel method for computing ambient occlusion from a stack of images of a Lambertian scene.
• To leverage AO for improved reflectance and illumination estimation without relying on additional smoothness priors.
• To demonstrate the method's state-of-the-art performance on established benchmarks and real-world data.

Main Methods:

• A simplified image formation model is used to approximate ambient occlusion via per-pixel statistics over image stacks.
• The derived AO measure is applied to compute object reflectance and illumination.
• The method processes image stacks from a fixed viewpoint for Lambertian scenes.

Main Results:

• The method successfully computes ambient occlusion from image stacks, approximating local visibility.
• Accurate reflectance and illumination were computed without additional smoothness priors.
• State-of-the-art performance was achieved on the MIT Intrinsic Images benchmark.
• The method was validated on synthetic and real scenes, including 3D printed objects with ground truth geometry.

Conclusions:

• The proposed method offers an effective way to compute ambient occlusion from image stacks, advancing computer vision applications.
• This approach enables robust reflectance and illumination estimation, outperforming existing methods on key benchmarks.
• The technique's versatility is demonstrated across various synthetic and real-world scenarios, highlighting its practical utility.