Abstract
There has been emerging interest on recovering traditionally challenging intrinsic scene properties. In this paper, we present a novel computational imaging solution for recovering the ambient occlusion (AO) map of an object. AO measures how much light from all different directions can reach a surface point without being blocked by self-occlusions. Previous approaches either require obtaining highly accurate surface geometry or acquiring a large number of images. We adopt a compressive sensing framework that captures the object under strategically coded lighting directions. We show that this incident illumination field exhibits some unique properties suitable for AO recovery: every ray's contribution to the visibility function is binary while their distribution for AO measurement is sparse. This enables a sparsity-prior based solution for iteratively recovering the surface normal, the surface albedo, and the visibility function from a small number of images. To physically implement the scheme, we construct an encodable directional light source using the light field probe. Experiments on synthetic and real scenes show that our approach is both reliable and accurate with significantly reduced size of input.
The Challenge of Ambient Occlusion
Ambient Occlusion (AO) is a critical component in realistic 3D rendering, providing the subtle shadows in crevices and folds that give objects depth. However, recovering AO from real-world images is notoriously difficult. Traditionally, researchers have had to choose between two suboptimal paths: performing a highly accurate 3D scan for ray-casting or capturing hundreds of images under varying light directions.
A significant limitation of previous image-based methods is their reliance on simplified parametric models, such as assuming visibility is a simple cone centered around the surface normal. As illustrated in Figure 1, real-world geometry often creates complex, non-parametric visibility functions—especially in "valleys" (point C) where light might be blocked from multiple directions.
Methodology: Compressive Visibility Estimation
This paper introduces a novel computational imaging approach that treats AO recovery as a Compressive Sensing (CS) problem. The core innovation lies in the observation that for any given surface point, the visibility of the hemisphere is a binary field—a ray is either blocked or it isn't—and this field is sparse in its distribution.
Coded Lighting and Hardware
To implement this, the authors use a "light field probe"—a microlens array placed in front of a projector—to create "coded" directional light. By using Hadamard patterns, the system can illuminate the object from multiple directions simultaneously. This increases the signal-to-noise ratio and allows the recovery of individual ray contributions from fewer measurements.
Iterative Solver
The authors formulate an optimization problem using $L_1$ and $L_\infty$ regularization to iteratively solve for three unknowns:
- Surface Albedo ($\rho$): The base reflectivity.
- Surface Normal ($n$): The orientation of the surface.
- Visibility Function ($V$): The binary map of occlusions.
The problem is split into two sub-problems: a visibility recovery step (solved as a second-order cone problem) and a normal recovery step (solved via constrained least squares).
Key Results
The method was validated using synthetic models (like the "Turtle" and "Tentacle") and real-world captures. The results demonstrate that the approach is both efficient and robust.
Efficiency and Accuracy
The proposed method achieves accuracy comparable to or better than previous techniques using only ~40 images, whereas prior methods like Hauagge et al. required 160 or more. As shown in Figure 4, the method significantly outperforms parametric models in concave regions where the "cone-shaped" visibility assumption breaks down. By recovering the full visibility function, the system provides cleaner albedo and AO maps in complex real-world scenes.
Conclusion and Future Work
By leveraging the sparse and binary nature of visibility, this framework provides a reliable way to recover AO maps without requiring perfect 3D geometry. While current limitations include the need for calibrated lighting and a Lambertian surface assumption, future work could incorporate spatial coherence (e.g., via Markov Random Fields) to further reduce noise and improve the smoothness of the recovered maps.
Summary updated