StereoSplat: Metric-Scale Novel View Synthesis via Stereo-Grounded Gaussian Splatting

StereoSplat is a feed-forward 3D Gaussian Splatting model built for stereo video, using features and disparity from a stereo backbone to predict per-pixel 3D Gaussians at metric scale. It adds a depth-adaptive aggregation step to reduce redundancy across overlapping views and was trained only on the new synthetic XRStereo dataset. The authors report improved depth accuracy in rendered novel views over baselines, with zero-shot transfer to real-world stereo video.