Graph-pruned semantic search

SemanticSplat

A hierarchical semantic graph that prunes scene search before grounding natural-language queries in captured evidence.

Read the paper Browse the code

Research prototype. Current claims are deterministic, test-backed, and scoped below.

97team-captured RGB-D views

1,066indexed semantic items

150same-input benchmark queries

75.5%fewer views checked

68.2%fewer estimated context tokens

Research question

Can hierarchy reduce query context without changing the semantic map?

SemanticSplat organizes the same semantic records used by flat search into scene, zone, region, object, and view-evidence levels. A query retains likely branches, ranks their leaves, and expands only when confidence or geometry requires fallback.

Safest current claim

Hierarchy substantially reduces checked views and serialized query context. The present lexical ranker does not yet preserve flat-search hit@k on the internal track, so the evidence supports a measured quality-cost tradeoff, not universal superiority.

Method

Captured evidence to graph-pruned grounding

The figure is assembled from the real Conference Hall RGB-D capture, its ViewJSON record, the checked-in hierarchy, and frozen query qv2_010. No AI-generated imagery is used.

SemanticSplat pipeline showing RGB-D capture, semantic ViewJSON, hierarchy construction, query parsing, graph-versus-flat traversal, and grounded evidence. — One-time construction is separated from per-query retrieval. The graph and flat lanes use identical records and scoring; only candidate selection changes.

Index evidence

Normalize objects, room cues, relations, optional 2D/3D boxes, and supporting views.

Build hierarchy

Attach deterministic entity records beneath scene, zone, and region branches.

Parse the query

Extract target, attributes, intent, optional anchor, and spatial relation.

Prune and rank

Gate branches, score retained entities, validate geometry, and expand on low confidence.

Hierarchy tree with retained green path and pruned gray branches. — Hierarchical search path: retained target evidence and pruned siblings.

Controlled same-map evaluation protocol comparing graph traversal with flat search. — Controlled protocol: same map, query, labels, and scorer.

Evidence

Results with protocol boundaries visible

Internal graph-versus-flat numbers are directly comparable. Public-dataset and external-baseline tracks answer different questions and are not collapsed into a single ranking.

Graph4.77views/query

Flat19.4views/query

Graph1,014est. context tokens/query

Flat3,168est. context tokens/query

Five captured scenes, 150 queries

Graph reduces mean views by 75.5% and estimated context tokens by 68.2%. On 125 queries with verified view labels, hit@1 is 0.680 graph versus 0.768 flat; hit@3 is 0.808 versus 0.928.

Inspect the frozen metrics JSON

Five-scene graph versus flat summary charts. — Efficiency and quality are reported together.

Execution	Scope	Measured status	Comparison boundary
ConceptGraphs / ScanNet	8 scenes, 48 queries	Native predicted-map run complete	Different map construction and query protocol
ConceptGraphs / Replica	8 scenes, 56 queries	Native predicted-map run complete	Different map construction and query protocol
LangSplat / ScanNet	1 scene, 6 queries	Full native pipeline complete	Hardware-adapted 3,000-iteration profile; not an 8-scene paper-scale reproduction

These executions establish integration and diagnostic reference points. They do not establish a fair method ranking against SemanticSplat.

Construction accounting

Measured graph-construction workload

The frozen audit separates one-time map preparation from per-query retrieval. It records the complete annotation payload, deterministic tree representation, serialized I/O, and clean-rebuild runtime for all five captured scenes.

Download construction audit

Manual records: 97 ViewJSON annotations
Semantic items: 1,066 manual payload entries
Annotation payload: 66,937 est. tokens; 267,877 chars
Constructed tree: 179,817 est. tokens
Serialized I/O: 246,754 input + output est. tokens
Local rebuild: 1.822 s sum of scene medians

Data

Exact evaluation sets and access boundaries

The page publishes scene IDs, manifests, derived metrics, and reproduction code. Licensed ScanNet and Replica scene files are not mirrored.

Team-ownedInternal track

Five manual captures

ConferenceHall-capture-pilot, Museume-capture, Theater-capture, outdoor-street-capture, and outdoor-drone-capture.

Views: 97
Items: 1,066
Queries: 150

Browse scene records

Terms requiredPublic pilot

ScanNet v2

scene0011_00, scene0030_00, scene0046_00, scene0086_00, scene0222_00, scene0378_00, scene0389_00, scene0435_00.

Scenes: 8
GT boxes: 392
Queries: 48

Request official ScanNet access

Official sourcePublic pilot

Replica

room0-room2 and office0-office4, aligned with the BBQ subset used in the paper.

Scenes: 8
GT boxes: 575
Queries: 56

Open the official Replica repository

Machine-readable scene lists and access notes: dataset manifest JSON.

Visual audit

Captured scenes and a grounded case

Gallery of six captured digital-twin scenes. — Team-captured internal scenes. Public-dataset imagery is intentionally not redistributed here.

Qualitative projection-screen grounding case with graph and flat views. — Qualitative query evidence with a view-annotation box.

Reproduce

Code, artifacts, and validation gates

Core verification

python -B -m pytest -q
python -B scripts/check_academic_paper.py
python -B scripts/measure_graph_construction.py --repeats 7
python -B scripts/export_academic_figures.py

Query implementationGrounding, traversal, provider modes Experiment scriptsBuild, run, adapt, evaluate, export Protocols and reportsGT, datasets, baselines, audits Canonical outputsMetrics and schema-valid query results

Limits

What this release does not claim

The five internal semantic indexes are manual reference annotations, not independent GT.
ScanNet and Replica SemanticSplat pilots use oracle semantic candidates; they do not measure semantic-perception accuracy.
ConceptGraphs and LangSplat use different map-construction and evaluation protocols, so their metrics are not a fair leaderboard.
Internal context tokens are deterministic text-size estimates. External baseline token counts are local tokenizer counts, not provider billing.
The current graph search saves context but trails flat lexical hit@k on the internal verified-view track.

Manuscript

SemanticSplat: Graph-Pruned Semantic Search

The public manuscript is a named-author academic preprint, independent of any workshop template. It includes the complete method, schemas, frozen metrics, uncertainty intervals, public-dataset pilots, external executions, and reproducibility contract.

Download PDF LaTeX source

Preprint citation

@article{mousatat2026semanticsplat,
  title = {SemanticSplat: Graph-Pruned Semantic Search},
  author = {Mousatat, Mahmoud and Vizan, Leo and
            Shankin, Nikita and Nuruzov, Telman and
            Medvedev, Alexandr},
  year = {2026},
  note = {Preprint}
}