Touch3D · Tac-DINO

Touch3D links GelSight Mini tactile contacts with object RGB, depth and 3D geometry. The Tac-DINO paper introduces CoMT-Bench for contact-centric multimodal evaluation and GeoTAC for tactile-centric cross-modal learning.

505 real-world objects · 20,025 tactile contacts · 18 material labels · 6 collection scenarios

Paper · Visual contact browser · All released objects

Manually labeled 3D contacts connect visible rendered views with tactile images and contact geometry. The right panel shows object-to-touch correspondences. Figure from the Tac-DINO manuscript; click to enlarge.

See the data

Inspect RGB, tactile, depth and geometry samples

Real data from one released object. RGB/depth use the same source view; tactile and contact geometry use contact 28, frame 10. Depth is colorized; point-cloud color encodes height.

Browse spatially annotated RGB–tactile pairs, contact locations and object metadata in Data Studio. Filter by category, material or stored object split. Preview rows retain source provenance.

Download

Selection	Contents	Compressed size
core	Rendered views, tactile clips and annotations	16.25 GB
geometry	Per-contact 3D geometry	57.05 GB
raw	Original scanner, video and depth assets	160.69 GB

pip install "huggingface_hub>=0.32,<3" "zstandard>=0.23"
curl -L https://huggingface.co/datasets/lihong-cs/Tac-DINO/resolve/main/download_dataset.py -o download_dataset.py
python download_dataset.py --group Scan3dTac --profile core --output ./data/scan3dtac

Use --object wipes/obj_0000 for one object. Add --profile geometry for per-contact 3D data, or use --group raw for originals. Core restoration needs about 55 GB of disk space.

from datasets import load_dataset

preview = load_dataset("lihong-cs/Tac-DINO", "contact_preview", split="samples", streaming=True)
sample = next(iter(preview))  # PIL images: sample["image"], sample["tactile_image"]

Data format and loading notes.

License and citation

CC BY 4.0 permits commercial use under its terms, including attribution, a license link and an indication of changes.

Please cite Tac-DINO: Learning Tactile Features with Patch Alignment, arXiv:2606.12069.