A Multi-View Photometric Stereo Pipeline for Specular 3D Fruit Reconstruction

Ariel Zuniga-Santana, Gabriele Facciolo, Shohei Nobuhara, Rodrigo Verschae; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 520-529

Reconstructing small fruits is challenging for traditional Multi-View Stereo (MVS) and classical photometric stereo because of their low texture, strong specularities, and limited feature visibility. Classical photometric stereo often fails on glossy surfaces, resulting in noisy normals. We introduce Agri-MVPS, which uses a polarized HDR photometric-normal reconstruction pipeline that overcomes these limitations and enables accurate geometry recovery for dark and glossy fruits. Then, Agri-MVPS integrates the reconstructed normals with neural implicit models, producing more complete and stable reconstructions than MVS or image-based methods alone and providing a foundation for 3D fruit datasets in precision agriculture.