Multi-view stereo and planar Gaussian splatting for generalizable surface reconstruction

Image-based three-dimensional (3D) reconstruction and novel view synthesis (NVS) play a critical role in applications such as virtual/augmented reality, 3D content generation, and autonomous driving. Recent studies have explored the integration of multi-view stereo (MVS) priors with 3D Gaussian splatting to improve cross-scene NVS. However, significant challenges remain in recovering fine rendering details and achieving accurate geometric surface reconstruction. To address these limitations, we propose a generalizable surface reconstruction method combining MVS with planar Gaussian splatting. Specifically, MVS geometric priors are employed to provide scene-adaptive initialization for Gaussian modeling, thereby enhancing global consistency and stabilizing cross-view rendering. In addition, we design a hybrid feature modeling architecture that integrates convolutional neural network, Mamba, and deformable convolutional network modules to jointly capture local details and long-range dependencies, leading to more robust multi-view feature representations. Furthermore, we introduce planar Gaussian modeling with unbiased depth estimation and geometric regularization, which improves surface continuity and overall reconstruction quality. Extensive experiments conducted on public benchmarks demonstrate that our method significantly outperforms existing generalizable methods in cross-scene NVS. Moreover, with per-scene optimization, our method achieves geometric accuracy and rendering quality that match or even surpass the latest state-of-the-art methods. This work provides a new perspective for developing 3D reconstruction methods that effectively balance efficient rendering, accurate surface recovery, and strong cross-scene generalization. The code is publicly available at https://github.com/yangyongjuan/MVSPG-SR .