Academic Journal
Adaptive spherical-harmonic allocation and saliency pruning for memory-efficient 3D Gaussian splatting.
| Τίτλος: | Adaptive spherical-harmonic allocation and saliency pruning for memory-efficient 3D Gaussian splatting. |
|---|---|
| Συγγραφείς: | Che, Xiaofeng1 (AUTHOR) binarystar671@foxmail.com, Wang, Chen1 (AUTHOR) cwang@cdut.edu.cn, Yao, Guangle1 (AUTHOR) yaoguangle19@cdut.edu.cn, Zhou, Haoran2 (AUTHOR) hr@sensing-sz.com, Ye, Shaoze2 (AUTHOR) 372767021@qq.com, Liu, Xulong3 (AUTHOR) liu.xulong@jingtian.com |
| Πηγή: | Visual Computer. Sep2026, Vol. 42 Issue 12, p1-19. 19p. |
| Θεματικοί όροι: | *Spherical harmonics, *Computer memory management, *Image quality in imaging systems |
| Περίληψη: | 3D Gaussian splatting (3DGS) has become a popular method for efficient and high-quality novel view synthesis. However, its training process demands excessive GPU memory, primarily due to uniform storage of high-order spherical-harmonic coefficients and rapid increases in the number of Gaussians. To address this, we introduce a memory-efficient approach that dynamically adjusts spherical-harmonic orders based on the expressive needs of each Gaussian and employs a saliency-based pruning strategy to remove redundant Gaussians. Our method reduces the peak GPU memory consumption during training by 29.6%, 29.8%, and 22.4% on the Mip-NeRF360, Tanks&Temples, and Deep Blending datasets, respectively, while maintaining rendering quality comparable to the original approach. [ABSTRACT FROM AUTHOR] |
| Βάση Δεδομένων: | Academic Search Index |
καταχωρήστε σχόλιο πρώτοι!