Accurate Normal Measurement of Non-Lambertian Complex Surface Based on Photometric Stereo
Yuze Yang, Jiahang Liu, Yue Ni, C. Li, Zihan Wang · IEEE Transactions on Instrumentation and Measurement · 2023
Photometric stereo network is a popular method for recovering the 3D surface of objects. Existing methods, both pixel- or spatial-based, often show poor performance in predicting surface normal in regions with sharp changes in concavity and convexity. In this paper, a pixel- and spatial-wised multi-feature fusion photometric stereo network (PSMF-PSN), which effectively combines inter- and intra-image information, is proposed to improve the deficiencies mentioned above. Firstly, normalized images and mask images are used to eliminate the adverse effects caused by discontinuities of surface material and by invalid background regions respectively, so as to improve the estimation accuracy of the surface normal. To extract features of pixels and their domains, a pixel-wised 3D feature extraction module and a spatial-wised feature extraction module are proposed and connected in a tandem manner. Furthermore, to capture more details, skip connections and max-pooling are used to fuse local-global and shallow-deep features. Experiments on DiLiGenT and DiLiGenT102benchmark datasets indicate that our proposed method achieves higher performance than the other photometric stereo methods. The qualitative tests on the Light Stage Data Gallery validate the effectiveness and generalizability of our method.