VidTouch

VidTouch is a specimen-level visuo-tactile benchmark for material understanding and material knowledge discovery. It links repeated RGB observations and DIGIT tactile videos of the same physical fabric to four semantic axes:

1. weave;
2. ordered material composition;
3. usage;
4. functional features.

We purchased 144 physical fabric swatches and independently acquired all released RGB images and tactile videos. We used the supplier descriptions as source records. As they were not provided as ready-to-use dataset labels, we extracted the relevant semantic information, translated it into English, designed the four-axis annotation schema covering weave, material, usage, and features, and constructed the initial annotation file. We then canonicalized spelling, separators, and genuine aliases while preserving meaning-bearing distinctions such as material composition order.

Release Contents

Item	Count
Fabric IDs	144
Usable RGB images	435
Usable DIGIT tactile videos	432
Complete weave labels	39
Complete material labels	60
Complete usage labels	46
Complete feature labels	100
Frozen benchmark labels	10 / 13 / 11 / 14
Frozen split	100 train / 22 validation / 22 test

Every retained Fabric ID has three tactile trials. Of the 144 fabrics, 142 have three RGB views, one has four, and one has five. RGB observations and tactile videos are independent repeated measurements grouped by Fabric ID; they are not frame-synchronized pairs.

VidTouch/
  RGBs/                         # 435 author-acquired JPEG images
    metadata.csv                # per-image labels and frozen split
  TACs/                         # 432 author-acquired MP4 tactile videos
    metadata.csv                # per-video labels and frozen split
  metadata/
    fabrics.csv                 # one row per Fabric ID
    observations.csv            # one row per released media file
  splits/
    fabric_common_v2.json       # frozen 100/22/22 benchmark split
    fabric_common_v2_lowshot25.json
    fabric_common_v2_lowshot50.json
  label.txt                     # complete canonical annotations
  release_manifest.json         # release statistics and excluded raw files
  checksums.sha256              # SHA-256 for every released file
  scripts/verify_release.py

Three Data Layers

Please distinguish the following layers when reporting results.

Usable media release. This is the complete 144-ID inventory of 435 RGB images and 432 tactile videos published in this repository.

Complete canonical annotations. label.txt conservatively normalizes spelling, separators, capitalization, and genuine aliases. It preserves supplier-provided Fabric IDs and preserves the component order of material blends because the order can represent descending composition. The complete vocabulary contains 39 weave, 60 material, 46 usage, and 100 feature labels.

Frozen common-label benchmark. Headline closed-set recognition uses the fixed fabric_common_v2.json split. Labels supported by fewer than three fabrics for weave/material/usage, or fewer than five fabrics for features, are masked only for their own attribute. A fabric is not globally deleted because one of its attributes is rare. This projection retains 10 weave, 13 material, 11 usage, and 14 feature labels.

The benchmark partitions are disjoint by Fabric ID. Do not regenerate the partition from a training seed, tune on Test, or treat repeated observations of one fabric as independent split units.

Annotation Format

Each non-comment line of label.txt is whitespace-delimited:

FabricID weave material usage feature_1 feature_2 ...

Weave, material, and usage are single-label attributes. Features are multi-label. Material strings are ordered compositions and must not be sorted or merged merely because they contain the same components.

The CSV metadata serializes the feature list as a JSON array. The split column uses train, validation, or test.

Loading

Download the complete immutable snapshot:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="AQUILA-espresso/VidTouch",
    repo_type="dataset",
)

Load each modality with the folder builders:

from datasets import load_dataset

rgb = load_dataset("imagefolder", data_dir=f"{root}/RGBs", split="train")
tactile = load_dataset("videofolder", data_dir=f"{root}/TACs", split="train")

rgb_train = rgb.filter(lambda example: example["split"] == "train")
tactile_train = tactile.filter(lambda example: example["split"] == "train")

Here, the folder builders call their complete local table train by default. That builder name is not the frozen benchmark partition. Always select the benchmark partition using the metadata split column or the Fabric IDs in splits/fabric_common_v2.json.

Video decoding through Hugging Face Datasets requires compatible ffmpeg and torchcodec installations. Reading the MP4 paths directly does not require TorchCodec.

For the official MESA training and evaluation pipeline, use VidTouch_MESA. Its data loader expects this repository layout directly.

Verify a downloaded snapshot:

python scripts/verify_release.py .

Collection and Sensor Interpretation

RGB views capture fabric appearance under indoor illumination. Tactile videos record DIGIT sensor observations during sustained-contact lateral stroking. Most trials begin with the sensor already loaded. Normal load can vary during the stroke, but the release does not provide a separately calibrated pressing trajectory.

DIGIT pixels combine elastomer deformation, internal illumination, moving texture, and operator motion. They are dynamic sensor-space evidence, not certified measurements of force, friction, stiffness, thickness, durability, comfort, safety, or product quality.

Intended Uses

Appropriate uses include:

• specimen-disjoint material and fabric recognition;
• visuo-tactile representation learning and cross-modal retrieval;
• long-tail, low-shot, and missing-modality evaluation;
• texture and contact-dynamics research;
• reproducible comparison on the frozen benchmark protocol.

The dataset is not intended to certify textile composition or performance, make safety-critical decisions, identify suppliers, or replace calibrated physical testing. Usage and functional-feature labels reflect supplier descriptions and may encode commercial terminology or annotation noise.

Integrity and Versioning

The canonical release excludes five RGB files and eleven tactile files from our raw collection because their Fabric IDs lack retained canonical annotations. Their names are recorded in release_manifest.jsonand the excluded media are not published as benchmark samples.

checksums.sha256 covers every released file. The frozen split also records its own split SHA-256, assignment SHA-256, and data-manifest SHA-256. Use a specific Hub revision when exact reproducibility matters.

Citation

The accompanying manuscript is under review. Until archival publication metadata is available, please cite the dataset repository:

@dataset{wang2026vidtouch,
  title     = {VidTouch: A Multimodal Physical Evidence Benchmark for Material Understanding},
  author    = {Wang, Yifan and Fu, Yu and Wang, Jie and Liu, Jinduo and Ma, Fenglong},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AQUILA-espresso/VidTouch}
}

License

The dataset is released under the license identified in the repository metadata and LICENSE file.