ARKitScenes-SpatialLM Dataset

ARkitScenes dataset preprocessed in SpatialLM format for oriented object bouding boxes detection with LLMs.

Overview

This dataset is derived from ARKitScenes 5,047 real-world indoor scenes captured using Apple's ARKit framework, preprocessed and formatted specifically for SpatialLM training.

Data Extraction

Point clouds and layouts are compressed in zip files. To extract the files, run the following script:

cd arkitscenes-spatiallm
chmod +x extract.sh
./extract.sh

Dataset Strucutre

arkitscenes-spatiallm/
├── arkitscenes_train.json # Training conversations
├── arkitscenes_val.json # Validation conversations
├── dataset_info.json # Dataset metadata
├── split.csv # Train/val split mapping
├── pcd/ # Point cloud data
│ └── .ply
├── layout/ # Scene layout annotations
│ └── .txt
└── extract.sh # Extraction script

The arkitscenes_train.json and arkitscenes_val.json dataset follows the SpatialLM format with ShareGPT-style conversations:

{
  "conversations": [
    {
      "from": "human",
      "value": "<point_cloud>Detect boxes. The reference code is as followed: ..."
    },
    {
      "from": "gpt",
      "value": "<|layout_s|>bbox_0=...<|layout_e|>"
    }
  ],
  "point_clouds": ["pcd/scene_id.ply"]
}

License

This dataset is derived from Apple's ARKitScenes dataset. Please refer to the original dataset's license terms for usage restrictions.

Citation

If you use this dataset in your research, please cite the original ARKitScenes paper:

@inproceedings{ARKitScenes,
  author    = {Gilad Baruch and Zhuoyuan Chen and Afshin Dehghan and Tal Dimry and Yuri Feigin and Peter Fu and Thomas Gebauer and Brandon Joffe and Daniel Kurz and Arik Schwartz and Elad Shulman},
  title     = {{ARKitScenes}: A Diverse Real-World Dataset for {3D} Indoor Scene Understanding using Mobile {RGB-D} Data},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year      = {2021}
}