StereoSplat

Abstract

Reconstructing three-dimensional (3D) scenes from a few input views remains difficult when dense image acquisition is impractical. Recent sparse-view surface reconstruction methods based on 3D Gaussian Splatting rely on stereo matching to guide optimization; however, the dense metric depth they compute is underexploited, serving merely as a loss signal that adjusts existing primitives. Therefore, this study proposes StereoSplat, which converts this overlooked depth into explicit scene geometry and thereby improves novel-view synthesis while preserving or enhancing surface reconstruction. Specifically, Stereo-depth Point Unprojection lifts dense stereo-depth maps into 3D seed points and injects them as new Gaussian primitives, filling coverage gaps left by the initial dense multi-view stereo point cloud in textureless and occluded regions. To ensure reliability, Stereo Quality Mask Selection retains only points that pass a left-right consistency check. Our analysis further shows that this within-view criterion is preferable to cross-view filtering, as the latter discards genuine surfaces visible from only a single view. In addition, Virtual View Photometric Augmentation enriches supervision by synthesizing virtual viewpoints near the training cameras and comparing the rendered appearance with forward-warped images. On the DTU dataset, StereoSplat attains a PSNR of 22.78 dB on the novel-view synthesis benchmark, surpassing state-of-the-art methods with consistent gains in SSIM, LPIPS, and AVGE. Moreover, surface accuracy is preserved under the large-overlap setting and improved under the little-overlap setting. Overall, these results reveal that initialization coverage, instead of optimization regularization, is the dominant bottleneck in Gaussian-based sparse-view reconstruction.