---
task_categories:
- zero-shot-object-detection
---

# Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding

This repository contains the ScanNet dataset (3D scene data and 2D frame data) and refined annotations used for the paper [Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding](https://huggingface.co/papers/2604.00528).

TAB is a dynamic agentic framework designed for zero-shot 3D Visual Grounding (3D-VG). By operating directly on raw RGB-D streams, TAB reformulates 3D grounding from a static proposal matching task into an active semantic reasoning and geometric reconstruction process.

- **GitHub:** [https://github.com/WHB139426/TAB-Agent](https://github.com/WHB139426/TAB-Agent)
- **Paper:** [https://huggingface.co/papers/2604.00528](https://huggingface.co/papers/2604.00528)

## ScanNet Data Preparation

This repository contains the ScanNet dataset split into two main compressed archives: 3D scene data and 2D frame data. 

### 1. Unzip the Data
Before using the data, please download and extract the zip files. You can use the following commands in your terminal:

```bash
# Unzip the 3D meshes and annotations (10.4 GB)
unzip scannet-dataset.zip

# Unzip the 2D RGB-D frames and camera poses (99.1 GB)
unzip scannet-frames.zip
```

### 2. Directory Structure

After unzipping, you will have two main directories organized by `scene_id`. 

#### `scannet-dataset/` (3D Meshes & Annotations)
This folder contains the core 3D spatial data, including the reconstructed meshes, semantic labels, and instance aggregations.

```text
scannet-dataset/
├── scene0000_00/
│   ├── scene0000_00_vh_clean_2.ply               # Cleaned 3D mesh
│   ├── scene0000_00_vh_clean_2.labels.ply        # 3D mesh with semantic labels
│   ├── scene0000_00_vh_clean_2.0.010000.segs.json # Over-segmentation file
│   ├── scene0000_00.aggregation.json             # Instance aggregation data
│   └── scene0000_00.txt                          # Scene metadata and properties
├── scene0000_01/
└── ...
```

#### `scannet-frames/` (2D Frames & Poses)
This folder contains the extracted 2D multi-view data for each scene, which is essential for tasks requiring RGB, depth, or camera pose alignments.

```text
scannet-frames/
├── scene0000_00/
│   ├── 00000.jpg   # RGB color frame
│   ├── 00000.png   # Depth map / Mask
│   ├── 00000.txt   # Camera pose matrix for this specific frame
│   ├── 00001.jpg
│   ├── 00001.png
│   ├── 00001.txt
│   └── ...
├── scene0000_01/
└── ...
```

## Citation

If you find this work useful, please consider citing:

```bibtex
@article{wang2026think,
  title={Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding},
  author={Wang, Haibo and Lin, Zihao and Xu, Zhiyang and Huang, Lifu},
  journal={arXiv preprint arXiv:2604.00528},
  year={2026}
}
```