Confidence-Aware Photometric Stereo Networks Enabling End-to-End Normal and Depth Estimation for Smart Metrology

The acquisition of geometric 3-D information is crucial for ensuring quality standards and monitoring procedures in various manufacturing applications. Photometric stereo is an established technique in computer vision to recover 3-D surfaces of objects. However, existing photometric stereo methods mainly focus on normal estimation of objects, without considering the depth estimation. On the other hand, current methods tend to prioritize accuracy while overlooking the confidence of predictions, which holds valuable information within the industry. In this article, we propose a deep learning-based photometric stereo system, consisting of hardware implementation, dataset generation, and algorithm design, to reconstruct 3-D information of physical objects, represented by normal and depth maps. In terms of the proposed algorithm, a coarse-to-fine network is introduced to improve the performance by exploiting the relationship between initial normal and depth predictions. Furthermore, the pixel-wise confidence associated with predictions is also estimated without requiring the ground truth, making a contribution to enhancing both performance and practicality. The experimental results on our synthetic dataset and real samples demonstrate the effectiveness of the proposed method on both normal/depth and confidence estimation.