Discrete Search Photometric Stereo for Fast and Accurate Shape Estimation.

Area of Science:

• Computer Vision
• Computer Graphics
• Photometric Stereo

Background:

• Estimating surface normals is crucial for 3D scene understanding.
• Existing methods often struggle with spatially varying Bidirectional Reflectance Distribution Functions (BRDFs) and computationally intensive optimization.
• Static camera observations under varying distant illuminations present challenges.

Purpose of the Study:

• To develop a novel, efficient method for surface normal estimation.
• To address limitations of continuous optimization in photometric stereo.
• To accelerate inference for scenes with complex BRDFs.

Main Methods:

• The problem is reformulated as a discrete search over discretized surface normals.
• Precomputation of scene-independent, expensive processes enables faster inference.
• Two variants of Discrete Search Photometric Stereo (DSPS) are proposed: one using continuous linear combinations of BRDF bases and another using discrete BRDF samples.

Main Results:

• DSPS achieves comparable accuracy to state-of-the-art exemplar-based methods.
• The proposed method demonstrates significant acceleration, achieving 10-100x speedup.
• Precomputation strategy effectively reduces inference time.

Conclusions:

• Discrete Search Photometric Stereo (DSPS) offers an efficient and accurate solution for surface normal estimation.
• The discrete search approach combined with precomputation is a viable alternative to continuous optimization.
• DSPS advances the field of photometric stereo by improving speed without sacrificing accuracy.