WACV 2027

RelateGS:
Reliable Semantic and Relational Evidence
for 3D Scene Understanding
with Gaussian Splatting

A structured language-aware Gaussian representation that selectively refines the evidence needed for relation-aware 3D visual grounding.

Seongjun Choi Sangyoun Lee^†

Yonsei University · School of Electrical and Electronic Engineering

^†Corresponding author

Paper (TBD)Code BibTeX

Figure 1. RelateGS filters unreliable semantic observations and complements under-observed relations using targeted re-querying of informative existing views.

Abstract

Grounding depends on the evidence behind a 3D scene.

Language-aware 3D Gaussian Splatting has advanced language-driven 3D scene understanding by incorporating semantic and relational cues into reconstructed scenes. However, relation-aware grounding depends not only on representing this information, but also on whether the underlying scene-side evidence is reliable and sufficient. RelateGS suppresses unreliable observations during semantic feature uplifting, consolidates multi-view object observations into a persistent 3D graph, and uses the graph to identify under-observed relations for targeted re-querying from informative existing frames. On Sr3D+ and Nr3D, RelateGS improves the previous best Acc@0.25 by 7.9 and 1.7 percentage points.

Method

Selective evidence refinement for a persistent 3D graph.

RelateGS improves the scene-side semantic and relational evidence supplied to grounding, rather than changing the downstream graph-matching protocol.

Figure 2. RelateGS first constructs reliable Gaussian semantics, then consolidates objects and relations into a persistent graph before resolving under-observed object pairs.

Confidence-aware semantic uplifting

Gaussian language features are aggregated from their supporting views, while observations with weak rendering significance are removed to suppress ambiguity from occlusion and limited visibility.

Persistent object-level relational evidence

Frame-level VLM observations are lifted into 3D and associated across views using spatial, semantic, and scale consistency, yielding persistent objects and accumulated relation evidence.

Graph-guided relational re-querying

Nearby object pairs with missing or low-confidence relations are located in the graph; an informative original camera view is retrieved and queried specifically for the deficient relation.

Evaluation

One representation, relation-aware scene understanding.

The paper evaluates both the primary grounding task and the semantic evidence underlying the persistent graph.

◎

3D Visual Grounding

Relation-aware references from Sr3D+ and Nr3D are resolved with the refined object graph and the same downstream matcher used by prior work.

◌

Semantic Segmentation

Gaussian-level language evidence is evaluated on Replica and ScanNet with mIoU and mean class accuracy.

⌕

Open-Vocabulary Localization

Language-aligned Gaussian features are evaluated on the four LERF-OVS scenes using Acc@0.25, Acc@0.5, and mIoU.

△

Controlled Ablations

Semantic observation selection and relational refinement are studied to isolate the contribution of each evidence source.

Qualitative results. Relation-aware 3D visual grounding comparison on ScanNet.

Results

Stronger grounding from refined scene-side evidence.

RelateGS achieves state-of-the-art overall 3D grounding performance at Acc@0.25 while preserving the same LLM-based graph matching protocol as GaussianGraph.

Relation-aware 3D visual grounding

Benchmark	Method	Easy A@0.25	Hard A@0.25	Overall A@0.25
Sr3D+	GaussianGraph	7.7	5.6	7.4
Sr3D+	RelateGS	15.2	6.6	15.3
Nr3D	GaussianGraph	12.1	6.3	8.6
Nr3D	RelateGS	13.2	6.4	10.3

Acc@0.25 (%). Full Easy, Hard, and overall Acc@0.1/0.25 results are reported in the manuscript.

Semantic segmentation

Dataset	Method	mIoU	mAcc
Replica	Occam's LGS	40.21	59.97
Replica	RelateGS	41.11	61.51
ScanNet	GaussianGraph	31.09	48.91
ScanNet	RelateGS	32.70	49.04

Citation

BibTeX

@inproceedings{choi2027relategs,
  title     = {RelateGS: Reliable Semantic and Relational Evidence for 3D Scene Understanding with Gaussian Splatting},
  author    = {Choi, Seongjun and Lee, Sangyoun},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year      = {2027}
}