README.md

metadata

license: cc-by-nc-4.0
pretty_name: R3D-Bench
configs:
  - config_name: qa_annotations
    default: true
    data_files:
      - split: test
        path: data/qa_annotations/*.parquet
  - config_name: gt_bboxes
    data_files:
      - split: test
        path: data/gt_bboxes/*.parquet
  - config_name: segmentations
    data_files:
      - split: test
        path: data/segmentations/*.parquet
  - config_name: meshes
    data_files:
      - split: test
        path: data/meshes/*.parquet

R3D-Bench

This is the official R3D-Bench dataset from R3D: Quantitative 3D Spatial Reasoning for Egocentric Wearables. R3D-Bench is a benchmark for quantitative 3D spatial reasoning from natural egocentric RGB-D video. 3,033 question-answer annotations over 57 Aria Digital Twin (ADT) sequences. This is an evaluation-only benchmark — there is no training split; all data is in the test split.

Note that project Aria video data must be downloaded separately. See full setup instructions in our github repository.

Configs

• qa_annotations (default): one row per question. Fields include question_text, gt_answer, question_type, sequence_id, spatial/temporal descriptions, and a nested referenced_objects list. This is the benchmark you evaluate against.
• segmentations: per-frame SAM3 masks (RLE) tracking relevant objects in the scene. Used in R3D-Bench evaluations for most methodologies. These are optional, but useful.
• meshes: per-object SAM3D reconstructed meshes with the GLB geometry embedded (glb bytes), used for evaluating the R3D methodology on R3D-Bench. They are available for use with other methodologies.
• gt_bboxes: optional. Ground-truth oriented-box trajectories from ADT that the reference answers were derived from. These are mainly for debugging. Evaluations do not use them.

Usage

from datasets import load_dataset
qa = load_dataset("facebook/r3d-bench", "qa_annotations", split="test")

Source video

The RGB-D frames are not redistributed here. Obtain the ADT sequences from https://www.projectaria.com/datasets/adt/ and extract frames as described in the R3D code repo.

Citation

When using our work, please cite as follows:

@misc{horton2026r3dquantitative3dspatial,
      title={R3D: Quantitative 3D Spatial Reasoning for Egocentric Wearables},
      author={Maxwell Horton and Wei Lu and Quan Tran and Yury Astashonok and Kirmani Ahmed and Babak Damavandi and Anuj Kumar and Xiao Zhang and Seungwhan Moon},
      year={2026},
      eprint={2607.02921},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.02921},
}

License

Derived from the Aria Digital Twin (ADT) dataset; use is subject to the ADT license/terms.