Photometric Stereo recovers the 3D shape of a surface from multiple images under different lighting directions. The reconstruction is formulated as a variational problem that reduces to solving the Poisson equation, which can be addressed with classical methods such as Jacobi, Gauss–Seidel, SOR, and multigrid. Simple methods reduce high-frequency errors but struggle with low-frequency ones, whereas multigrid handles both efficiently. I present a UNet–based model trained only on synthetic data as a fast Poisson solver for face surface reconstruction, achieving accuracy comparable to multigrid while being significantly faster on real face datasets.
