Ziyannn/sequential-3d-grounding

Sequential 3D Visual Grounding Dataset

5 indoor datasets (ScanNet, HM3D, 3RScan, ARKitScenes, MultiScan) · 10301 scenes · 117885 sequences · 579872 steps

Overview

Dataset	Scenes	Train	Val	Test
ScanNet	1513	14501	1700	1808
HM3D	2302	33302	7663	4232
3RScan	1381	13826	4864	1937
ARKitScenes	4834	24349	1394	2653
MultiScan	271	4133	858	665
Total	10301	90111	16479	11295

Annotation

Each step has one target (the object to locate) and zero or more anchors (reference objects). Anchors are extracted from prepositional objects, verified against scene instances, and enforced different from target.

• target_id non-null: 99.98%
• Anchor distribution: 0 = 1.8%, 1 = 36.0%, ≥2 = 62.2%
• 0 out-of-range / anchor==target violations

Full spec: DATASET_RECONSTRUCTION_GUIDE.md (English + 中文)

Structure

├── annotations/
│   ├── dynamic_data/          # {dataset}_{train,val,test}.json
│   ├── splits/                # {dataset}_{train,val,test}.txt
│   └── meta_data/             # 607 categories, synonym map, label TSV
├── scan_data/                 # 3D-VisTA instance attributes
│   └── {Dataset}/scan_data/
│       ├── instance_id_to_name/    # {scene}.json (list[str])
│       ├── instance_id_to_loc/      # {scene}.npy ((n,6) center+size)
│       └── instance_id_to_gmm_color/ # {scene}.json (color GMM)
└── final_qc_report.json

Usage

1. Point 3D-VisTA's SCAN_FAMILY_BASE to scan_data/.
2. Load annotations/dynamic_data/{dataset}_{split}.json.
3. Strip {Dataset}_ prefix from scan_id to get the scene name.

Source code and reproduction scripts: GitHub

Sources

• Point clouds: SceneVerse
• Sequential annotations: SG3D + SceneVerse SSG
• Target/anchor definition: v2 grounding-centric (this work)

License

Refer to the original dataset licenses (ScanNet, HM3D, 3RScan, ARKitScenes, MultiScan) and SceneVerse terms.