Texture-aware 3D Gaussian Splatting for sparse view reconstructions

Recently, 3D Gaussian Splatting (3DGS) has achieved high rendering quality in Novel View Synthesis (NVS). However, as the number of input views decreases, 3DGS fails to recover the details of the captured 3D scene due to insufficient constraints. We find that the difficulty of Gaussian primitives to concentrate on texture-rich areas leads to this reconstruction degradation. To this end, we propose TA-GS, a texture-aware framework for sparse-view NVS tasks. Specifically, TA-GS introduces a Texture-Based Gaussian Migration strategy, which detects low-opacity Gaussian primitives and migrates them to texture-rich regions, improving the fidelity of texture representation. Additionally, we utilize the texture of depth maps and introduce a Depth Texture Alignment method to constrain the geometric structures. To prevent overfitting to sparse input views, TA-GS employs Phantom View Regularization to enrich texture information from additional phantom views. Extensive experiments demonstrate that our approach outperforms previous methods across a variety of datasets, including LLFF, Mip-NeRF360, DTU, and Blender.