Generalizable 3D Gaussian splatting via multi-view stereo and consistency constraints

Recent neural rendering methods still struggle with fine-grained detail reconstruction and scene generalization, especially when handling complex geometries and low-texture regions. To address these challenges, we propose a 3D Gaussian Splatting (3DGS) framework enhanced by Multi-view Stereo (MVS), aiming to improve both rendering quality and cross-scene adaptability. Specifically, we first introduce an Adaptive Perception-aware Feature Aggregation (APFA) module, which effectively fuses 2D image features into 3D geometry-aware representations via a Local Feature Adaptive Collaboration (LFAC) mechanism and a global Attention-Aware Module (AAM), significantly improving reconstruction performance in challenging scenes. Subsequently, we propose a depth and normal supervision strategy based on multi-view geometric consistency, where aggregated point clouds are utilized for optimized initialization, enhancing stability and fine-grained detail fidelity. Finally, a Gaussian geometric consistency regularization module is introduced to further enforce the coherence between depth and normal predictions, leading to more refined rendering results. Extensive experiments on standard benchmarks including DTU, Real Forward-facing, NeRF Synthetic, and Tanks and Temples demonstrate that our method outperforms state-of-the-art approaches in terms of PSNR, SSIM, and LPIPS metrics. Particularly in real-world complex scenes, our approach achieves superior generalization ability and perceptual quality, validating the effectiveness of the proposed framework. The code for our method will be made available at https://github.com/yangyongjuan/MVS-APFA-GS.