SparseIR: improving sparse-view inverse rendering via integrating monocular geometric and material priors

Inverse rendering involves recovering the shape, material, and illumination of a 3D scene from captured images. Recent approaches achieve high-quality scene element recovery leveraging scene representations. However, when training views are sparse, the ill-posed nature of inverse rendering is amplified, making it difficult for these methods to achieve satisfactory results. We present SparseIR, an efficient inverse rendering framework that integrates monocular priors into 3D Gaussian splatting to achieve efficient scene decomposition and relighting with sparse views. The priors include geometry (normal) and material (albedo) derived from pre-trained models. First, we apply the normal prior to constrain the smoothed geometry, thereby preserving geometric details. Subsequently, to address scaling discrepancies in the albedo prior, we design a loss function based on linear correlation to regulate the material. Extensive experiments on synthetic and real datasets demonstrate that our approach outperforms existing inverse rendering methods under sparse view conditions.