DATASET_RECONSTRUCTION_GUIDE.md

Sequential 3D Visual Grounding — Technical Documentation / 技术文档

Version v2 · 2026-09-20

---

English

This dataset targets Sequential 3D Visual Grounding in indoor scenes. It covers 5 datasets (ScanNet, HM3D, 3RScan, ARKitScenes, MultiScan), 10301 scenes, 117885 task sequences, and 579872 action steps. Data is converted to 3D-VisTA-compatible format.

1. Task Definition

Given a 3D scene (point cloud + instance segmentation) and an ordered sequence of natural language instructions, the model locates the described object at each step. Unlike single-turn grounding, adjacent steps share semantic context — objects may recur across steps, pronouns may refer to previously mentioned objects.

2. Target & Anchor Annotation

Concept	Definition	Constraint
target	The object the model must locate in the current step	Exactly 1 per step, never null
anchor	Reference object(s) that help disambiguate the target	0, 1, or more; must differ from target
anchor_type	Source of the anchor	explicit (in current sentence) / historical (inherited)

Key rules:

• Navigation steps ("walk to X"): destination X is the target.
• Action steps ("pick up Y on Z"): Y is target, Z is anchor.
• Anchors are extracted from prepositional objects only.
• anchor ≠ target enforced at the instance-id level.

3. Data Sources

Source	Content
SceneVerse	Unified scanfamily point clouds + instance labels
SG3D	Sequential grounding annotations
SceneVerse SSG	Chain-style referring expressions
This work	v2 target/anchor redefinition

All instance labels aligned to ScanNet's 607 categories (verified 100% hit rate).

4. Annotation Schema

{
  "scan_id": "ScanNet_scene0350_00",
  "action_steps": [
    {
      "action_id": 1,
      "action": "Walk to the sink underneath the soap dispenser.",
      "target_id": "14",
      "target_object": "sink",
      "target_changed": true,
      "anchors": [
        {"anchor_id": "24", "anchor_object": "soap dispenser", "anchor_type": "explicit"}
      ],
      "anchor_source": "rule_backfill"
    }
  ]
}

5. Statistics

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

Anchor distribution (per step): 0 = 1.8%, 1 = 36.0%, ≥2 = 62.2%.

6. Quality Control

Check	Result
target_id missing	104 (0.018%, 3RScan unresolvable)
target_id out of range	0
anchor_id out of range	0
anchor == target	0
anchor-target distance (median)	0.32× scene scale

7. 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.
4. Add an anchor head alongside the existing grounding head; train with multi-task loss L = L_target + λ·L_anchor.

8. Scripts

Script	Function
scanfamily_to_3dvista.py	Point cloud format conversion
resolve_target_anchors.py	LLM names → instance IDs
build_dynamic_sequences.py	SSG chain windowing + assembly
filtered_to_v2.py	Legacy data → v2 schema
backfill_filtered_anchors.py	Backfill missing prepositional anchors
llm_augment_anchors.py	Spatially-verified LLM augmentation (train only)
generate_unified_splits.py	Scene-level splits
final_qc.py	Exhaustive quality check

---

中文

本数据集面向 3D 室内场景中的连续指令定位（Sequential 3D Visual Grounding）研究。 覆盖 ScanNet、HM3D、3RScan、ARKitScenes、MultiScan 共 5 个数据集、10301 个场景， 117885 个任务序列、579872 个动作步。

1. 任务定义

给定 3D 场景（点云 + 实例分割）和一组有序自然语言指令序列， 模型在每一步定位该步所描述的物体。与单轮 grounding 的区别：相邻步之间存在语义关联， 同一物体可能跨步复现，代词可能指代前序步骤中已出现的对象。

2. Target 与 Anchor 标注

概念	定义	约束
target	当前步模型必须定位的物体	每步恰好 1 个，不缺失
anchor	帮助确定 target 位置或消歧的参照物体	0、1 或多个；与 target 不同
anchor_type	anchor 来源	explicit（当前句）/ historical（跨步继承）

判定规则：

• 导航步（walk to X）：目的地 X 为 target。
• 操作步（pick up Y on Z）：Y 为 target，Z 为 anchor。
• anchor 仅取介词宾语，排除非空间参照的普通名词。
• id 层强制 anchor ≠ target。

3. 数据来源

来源	内容
SceneVerse	五数据集统一点云与实例标签
SG3D	序列化 grounding 标注
SceneVerse SSG	chain 式指代表达
本数据集	v2 target/anchor 重定义

所有实例标签对齐 ScanNet 607 类（逐数据集验证 100% 命中）。

4. 标注格式

{
  "scan_id": "ScanNet_scene0350_00",
  "action_steps": [
    {
      "action_id": 1,
      "action": "Walk to the sink underneath the soap dispenser.",
      "target_id": "14",
      "target_object": "sink",
      "target_changed": true,
      "anchors": [
        {"anchor_id": "24", "anchor_object": "soap dispenser", "anchor_type": "explicit"}
      ],
      "anchor_source": "rule_backfill"
    }
  ]
}

5. 统计

数据集	Train	Val	Test
ScanNet	14501	1700	1808
HM3D	33302	7663	4232
3RScan	13826	4864	1937
ARKitScenes	24349	1394	2653
MultiScan	4133	858	665
合计	90111	16479	11295

Anchor 分布（每步）：0 个 = 1.8%，1 个 = 36.0%，≥2 个 = 62.2%。

6. 质检

检查项	结果
target_id 缺失	104 (0.018%，3RScan 无法解析)
target_id 越界	0
anchor_id 越界	0
anchor == target	0
anchor-target 距离中位数	0.32 倍场景尺度

7. 使用方法

1. 将 3D-VisTA 的 SCAN_FAMILY_BASE 指向 scan_data/。
2. 加载 annotations/dynamic_data/{dataset}_{split}.json。
3. scan_id 去掉 {Dataset}_ 前缀得到场景名。
4. 在现有 grounding head 旁增加 anchor head，多任务损失 L = L_target + λ·L_anchor。

8. 脚本

脚本	功能
scanfamily_to_3dvista.py	点云格式转换
resolve_target_anchors.py	LLM 名词 → 实例 id
build_dynamic_sequences.py	SSG chains 切窗 + 组装
filtered_to_v2.py	旧标注转 v2
backfill_filtered_anchors.py	介词参照漏标补齐
llm_augment_anchors.py	空间校验 LLM 增强（仅 train）
generate_unified_splits.py	场景级切分
final_qc.py	全面质检

---

Machine-readable QC report: final_qc_report.json