PAGaS: Pixel-Aligned 1DoF Gaussian Splatting for Depth Refinement

Researchers from Meta, TU Munich, and the University of Zaragoza (David Recasens, Robert Maier, Aljaz Bozic, Stephane Grabli, Javier Civera, Tony Tung, Edmond Boyer) have published a paper titled PAGaS: Pixel-Aligned 1DoF Gaussian Splatting for Depth Refinement. The paper introduces a new method that adapts Gaussian Splatting (GS) — originally designed for novel view synthesis — to the multi-view stereo depth estimation task. The team's key contribution is modeling each pixel's depth using one-degree-of-freedom (1DoF) Gaussians that remain tightly constrained during optimization, unlike previous GS variants that struggle with geometric accuracy.

PAGaS restricts each Gaussian's position and size to the back-projected volume of its corresponding pixel, leaving depth as the sole degree of freedom to optimize. This constraint dramatically improves geometric fidelity while maintaining the high rendering quality expected from GS-based methods. On challenging 3D reconstruction benchmarks, PAGaS outperforms both classical geometric and modern learning-based multi-view stereo baselines, producing highly detailed depth maps. The code is available on GitHub, enabling researchers and practitioners to reproduce results and integrate the approach into existing pipelines. The paper is currently under review and available on arXiv.