ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos
ClearGS weights imperfect handheld frames instead of throwing them out, then repairs usable detail without clean reference images.
The paper targets 3D Gaussian Splatting from handheld video where viewpoints are uneven and frames vary in quality. Its Reliability-aware View Allocation assigns graded supervision weights using appearance reliability, degradation risk, and geometric utility, while keeping some suppressed frames active for trajectory coverage. A second stage, Render-Guided In-Video Restoration, uses the current 3DGS render and a frozen no-reference restoration model to choose between rendered, restored, and fused candidates. The authors report state-of-the-art results on GS2E and GSOTM, including CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings. ArXiv · AI/CL/LG's note
The paper targets 3D Gaussian Splatting from handheld video where viewpoints are uneven and frames vary in quality. Its Reliability-aware View Allocation assigns graded supervision weights using appearance reliability, degradation risk, and geometric utility, while keeping some suppressed frames active for trajectory coverage. A second stage, Render-Guided In-Video Restoration, uses the current 3DGS render and a frozen no-reference restoration model to choose between rendered, restored, and fused candidates. The authors report state-of-the-art results on GS2E and GSOTM, including CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings. ArXiv · AI/CL/LG's note
score 4