BROWSE
Detail
Journal Article
Object-Aware Loss for Mask-Guided Mesh Reconstruction Based on 2D Gaussian Splatting
- Authors
- Sunghyun Woo, Jeongil Seo, Heekyung Lee
- Issue Date
- 2026-09
- Citation
- Mathematics, v.14, no.17, pp.2-32
- ISSN
- 2227-7390
- Publisher
- Multidisciplinary Digital Publishing Institute (MDPI)
- Language
- English
- Type
- Journal Article
- Abstract
- Gaussian Splatting has emerged as a powerful representation for high-fidelity novel view synthesis. However, existing methods typically optimize the entire scene and require postprocessing to obtain object-level meshes. This process incurs unnecessary computational costs and can introduce background artifacts into the resulting object mesh. To address this problem, we propose an object-aware loss for mask-guided mesh reconstruction based on 2D Gaussian Splatting, which combines three complementary terms. First, the Masked L1 Loss restricts photometric supervision to the target-object region. Second, the Alpha Polarization Loss encourages high accumulated alpha values on the object and low values on the background. Third, the ObjectMark Polarization Loss supervises a learnable ObjectMark score, and Gaussians with scores at or below a predefined threshold are periodically pruned during optimization. By suppressing background-contributing Gaussians and focusing optimization on the target object, the proposed method enables efficient object-level reconstruction with fewer background artifacts. Among the compared methods, the proposed method achieves the highest surface reconstruction accuracy. It reduces the Chamfer Distance on the DTU dataset from 0.92 to 0.74, a 19.6% improvement, while producing cleaner object-level meshes. These results show that the proposed method enables direct and efficient object-level mesh reconstruction while reducing the need for postprocessing.
- Keyword
- 3D reconstruction, Gaussian Splatting, mesh extraction, novel view synthesis, surface reconstruction
- KSP Keywords
- 2D Gaussian, 3d reconstruction, Chamfer distance, Focusing optimization, High-fidelity, Mesh extraction, Mesh reconstruction, Novel view synthesis, Object region, Object-aware, Object-level
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
