Multiview photometric stereo for dense 3D reconstruction using a light-attached camera

Photometric stereo in a dynamic environment presents challenges in both establishing pixel-wise correspondences of images and dealing with close-range lighting. To overcome these problems, we propose a multiview photometric stereo method for dense and complete 3D reconstruction, specifically tailored to perspective cameras and close-range lighting scenarios. Our method only relies on structure from motion algorithms with estimated camera motion and sparse point clouds of the target object. With these priors, we use LoFTR to obtain feature matching and then use the thin plate spline interpolation method to further establish robust pixel-wise correspondences and to correct reconstruction deviation caused by close-range lighting. Furthermore, to solve the problem of traditional photometric stereo only performing partial 2.5D reconstruction, we propose a photometric stereo multiview stitching method to achieve complete 3D reconstruction. Our data capture setup is simple, consisting of only a light source and a digital camera. Evaluations with real-world objects demonstrate that our method achieves high-quality, complete 3D reconstructions with detailed geometric accuracy.