{"title":"HR-2DGS: Hybrid regularization for sparse-view 3D reconstruction with 2D Gaussian splatting","authors":"Yong Tang, Jiawen Yan, Yu Li, Yu Liang, Feng Wang, Jing Zhao","doi":"10.1016/j.cag.2025.104444","DOIUrl":null,"url":null,"abstract":"<div><div>Sparse-view 3D reconstruction has garnered widespread attention due to its demand for high-quality reconstruction under low-sampling data conditions. Existing NeRF-based methods rely on dense views and substantial computational resources, while 3DGS is limited by multi-view inconsistency and insufficient geometric detail recovery, making it challenging to achieve ideal results in sparse-view scenarios. This paper introduces HR-2DGS, a novel hybrid regularization framework based on 2D Gaussian Splatting (2DGS), which significantly enhances multi-view consistency and geometric recovery by dynamically fusing monocular depth estimates with rendered depth maps, incorporating hybrid normal regularization techniques. To further refine local details, we introduce a per-pixel depth normalization that leverages each pixel’s neighborhood statistics to emphasize fine-scale geometric variations. Experimental results on the LLFF and DTU datasets demonstrate that HR-2DGS outperforms existing methods in terms of PSNR, SSIM, and LPIPS, while requiring only 2.5GB of memory and a few minutes of training time for efficient training and real-time rendering.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"133 ","pages":"Article 104444"},"PeriodicalIF":2.8000,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Computers & Graphics-Uk","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0097849325002857","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, SOFTWARE ENGINEERING","Score":null,"Total":0}
引用次数: 0
Abstract
Sparse-view 3D reconstruction has garnered widespread attention due to its demand for high-quality reconstruction under low-sampling data conditions. Existing NeRF-based methods rely on dense views and substantial computational resources, while 3DGS is limited by multi-view inconsistency and insufficient geometric detail recovery, making it challenging to achieve ideal results in sparse-view scenarios. This paper introduces HR-2DGS, a novel hybrid regularization framework based on 2D Gaussian Splatting (2DGS), which significantly enhances multi-view consistency and geometric recovery by dynamically fusing monocular depth estimates with rendered depth maps, incorporating hybrid normal regularization techniques. To further refine local details, we introduce a per-pixel depth normalization that leverages each pixel’s neighborhood statistics to emphasize fine-scale geometric variations. Experimental results on the LLFF and DTU datasets demonstrate that HR-2DGS outperforms existing methods in terms of PSNR, SSIM, and LPIPS, while requiring only 2.5GB of memory and a few minutes of training time for efficient training and real-time rendering.
期刊介绍:
Computers & Graphics is dedicated to disseminate information on research and applications of computer graphics (CG) techniques. The journal encourages articles on:
1. Research and applications of interactive computer graphics. We are particularly interested in novel interaction techniques and applications of CG to problem domains.
2. State-of-the-art papers on late-breaking, cutting-edge research on CG.
3. Information on innovative uses of graphics principles and technologies.
4. Tutorial papers on both teaching CG principles and innovative uses of CG in education.