The X-Capture dataset contains multisensory data collected from 600 real-world objects in nine in-the-wild environments. We provide RGB-D, acoustic, tactile, and 3D data. Each object has six recorded points each, covering diverse locations on the object.

The dataset can be downloaded with:

wget https://huggingface.co/datasets/swistreich/XCapture/resolve/main/XCapture_data.zip -O XCapture_data.zip

Usage

Cross-sensory retrieval (audio→image, touch→3D, etc.)
Multimodal representation learning
Pretraining encoders across RGB-D / tactile / audio
Object-centric perception
Reconstruction (2D/3D) from X-modal signals

Dataset Structure

Each object directory contains six capture points, each with:

rgb: 640×480 color images
depth: aligned depth images
tactile: high-resolution DIGIT tactile images under 10N, 15N, 20N presses
audio: ~3s audio/video clip of impact sound
3D: local object mesh at contact point

There are no train/val/test splits; users are encouraged to construct splits suited to their task.

Citation

BibTeX:

@misc{clarke2025xcapture,
    title={X-Capture: An Open-Source Portable Device for Multi-Sensory Learning},
    author={Samuel Clarke and Suzannah Wistreich and Yanjie Ze and Jiajun Wu},
    year={2025},
    eprint={2504.02318},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2504.02318},
}