Li, Zhida, Zhu, Jianqiao, Huang, Hejin, Qin, Yipeng  ORCID: https://orcid.org/0000-0002-1551-9126, Yang, Sibei and Li, Guanbin
2026.
Direct language embedding enables Gaussian splatting for large scenes.
Presented at: Findings of the Conference on Computer Vision and Pattern Recognition (CVPR Findings) 2026,
Denver, CO, USA,
3-7 June 2026.
Proceedings of CVPRF.
IEEE,
pp. 7234-7243.
10.1109/CVPRF71909.2026.00694
|
Preview |
PDF
- Accepted Post-Print Version
Download (9MB) | Preview |
Abstract
Language Embedded Gaussian Splatting (LEGS) offers a promising direction for 3D scene understanding, yet existing approaches rely heavily on dense geometric priors derived from costly 3D reconstruction, which often lead to inaccuracy and inefficiency when scaled to large and complex scenes. In this work, we address this limitation by introducing Direct Language Embedded Gaussian Splatting (DiLEGS): a novel framework that enhances both the effectiveness and efficiency of LEGS, enabling its scalable and practical deployment. Specifically, (i) we challenge the prevailing assumption that geometric priors obtained through costly 3D reconstruction are beneficial for scene understanding. Contrary to this belief, we show that such an assumption breaks down in large scenes, thereby proposing a reconstruction-free pipeline. (ii) To address the challenge of inconsistent 2D semantic masks in the absence of geometry priors, we introduce a novel temporal-spatial mask synchronization strategy for mask alignment that substantially enhances accuracy while incurring minimal computational overhead. (iii) Observing that the standard SfMinitialization of LEGS is overly dense for semantic modeling, we propose a sparse initialization trick that initializes LEGS with a much sparser set of Gaussian primitives, improving efficiency without compromising quality. Extensive experiments on large-scale datasets demonstrate the superiority of our method, paving the way for an effective and efficient solution to real-world 3D scene understanding.
| Item Type: | Conference or Workshop Item - published (Paper) |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Computer Science & Informatics |
| Publisher: | IEEE |
| ISBN: | 9798319547583 |
| Date of First Compliant Deposit: | 8 April 2026 |
| Date of Acceptance: | 21 February 2026 |
| Last Modified: | 19 Oct 2026 13:00 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/186160 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions