3DSAV: 3D Scene Assertion Verification

3DSAV is a benchmark for verifying natural-language assertions against 3D indoor scenes. Given a scene and an assertion, a model predicts a binary answer, yes or no. The dataset contains 22,520 annotations across 682 ScanNet scenes, covering six semantic types.

This repository hosts the assertion annotations and their original scene-disjoint splits. ScanNet geometry, RGB frames, and model predictions are not included.

Load the dataset

from datasets import load_dataset

dataset = load_dataset("csdjade/3DSAV")
print(dataset)
print(dataset["train"][0])

The validation split corresponds to 3dsav_val.json in the original release.

Splits

Split	Annotations	ScanNet scenes	yes	no
train	18,620	547	9,005	9,615
validation	2,001	68	924	1,077
test	1,899	67	875	1,024
Total	22,520	682	10,804	11,716

Question IDs and scene IDs are mutually disjoint across the three splits. The test split contains the original 1,899 examples.

Fields

Field	Type	Description
question_id	string	Unique annotation identifier.
scene_id	string	ScanNet scene identifier, such as scene0667_00.
q	string	Natural-language assertion or binary verification question.
a	string	Ground-truth answer: yes or no.
name	list of strings	Object names supplied with the annotation.
ids	list of integers	Corresponding object IDs in the source scene annotation.
type	string	One of the six semantic types below.

Object IDs are scene-local. The name and ids lists have matching lengths.

{
  "question_id": "scene0667_00-23",
  "scene_id": "scene0667_00",
  "q": "Is there a stool under the table?",
  "a": "yes",
  "name": ["table", "stool"],
  "ids": [5, 43],
  "type": "spatial"
}

Semantic types

Type	Description	Total
spatial	Spatial relationships between scene elements.	5,532
negation	Negative logic and its scope.	4,421
attribute	Observable visual properties of objects.	4,171
quantity	Counting and quantification.	3,746
existence	Whether an object is present.	2,576
verification	Comprehensive assertions outside the other five categories.	2,074

The benchmark was constructed through scene-aware language-model generation followed by human verification.

Release format

The three JSONL files preserve the original fields, records, labels, and row order. The type field was added to each split by an exact, unique join on question_id against 3dsav_all.json. Every original field was checked against that full annotation file.

3DSAV_original_annotations.zip contains the four original JSON files with their contents unchanged: 3dsav_all.json, 3dsav_train.json, 3dsav_val.json, and 3dsav_test.json.

Scene data and usage

Obtain the underlying scenes through the ScanNet project and follow its access conditions and terms. Use scene_id to connect an assertion to its scene. Evaluating the benchmark requires the corresponding scene data; the text annotations alone do not provide the visual evidence.

The source annotation package does not specify a separate annotation license. No additional license is asserted by this upload. ScanNet's terms continue to apply to the underlying scene data.

Reference

3D Scene Assertion Verification. Jun Lin, Jiayu Ding, Xiangtian Si, Xitong Cao, Lixin Hong, Zhang Chen, Chenxi Lv, and Wenqian Wang, 2026.