---

HUST-SOFT

Dataset Description

HUST-SOFT is a vision-based tactile dataset developed for studying tactile deformation, contact-force perception, geometric-feature perception, and material compliance recognition.

The dataset contains tactile deformation images collected during controlled interactions between a vision-based tactile sensor and silicone rubber specimens with different geometric shapes and compliance levels. The specimens were fabricated from silicone rubber polymers with varying cross-linking ratios to finely tune their inherent stiffness.

Each sample corresponds to one complete tactile interaction. A sample contains three tactile deformation images acquired at three different moments during the interaction, together with the corresponding z-axis contact force for each image.

HUST-SOFT contains:

• 901 tactile interaction samples;
• 2,703 tactile deformation images;
• 3 geometric shape categories;
• 13 Shore A hardness levels in total;
• 3 tactile deformation images per interaction;
• z-axis contact-force measurements associated with each image.

The geometric shape and Shore A hardness categories do not form a complete Cartesian combination. The angle and edge categories contain samples with Shore A hardness values of 31, 41, 53, 64, 73, 82, and 87. The circular category contains samples with Shore A hardness values of 27, 39, 49, 62, 71, and 79.

Dataset Structure

The dataset is organized hierarchically according to geometric shape, Shore A hardness, interaction index, and tactile deformation image:

HUST-SOFT/
├── angle/
│   ├── 31/
│   │   ├── 1/
│   │   │   ├── <force_value_1>.jpg
│   │   │   ├── <force_value_2>.jpg
│   │   │   └── <force_value_3>.jpg
│   │   ├── 2/
│   │   └── ...
│   ├── 41/
│   ├── 53/
│   ├── 64/
│   ├── 73/
│   ├── 82/
│   └── 87/
├── circular/
│   ├── 27/
│   ├── 39/
│   ├── 49/
│   ├── 62/
│   ├── 71/
│   └── 79/
└── edge/
    ├── 31/
    ├── 41/
    ├── 53/
    ├── 64/
    ├── 73/
    ├── 82/
    └── 87/

The complete file path follows the format:

HUST-SOFT/<shape>/<shore_a_hardness>/<interaction_index>/<force_value>.jpg

For example:

HUST-SOFT/angle/31/1/256.jpg

represents:

• angle: the geometric shape category;
• 31: a Shore A hardness value of 31;
• 1: the interaction index;
• 256.jpg: a tactile deformation image acquired when the corresponding z-axis contact force was 2.56 N.

Each interaction directory contains exactly three tactile deformation images from the same interaction process.

Image Format

All tactile deformation images have the following properties:

Property	Description
File format	JPEG
Image resolution	640 × 480 pixels
Frames per interaction	3
Associated measurement	Z-axis contact force
Force unit	N
Force encoding	Numerical filename divided by 100

File-Naming Convention

The numerical part of each JPEG filename is 100 times the corresponding z-axis contact force in newtons.

The contact force can therefore be recovered using:

contact force in N = numerical filename / 100

Examples are provided below:

Filename	Z-axis contact force
85.jpg	0.85 N
256.jpg	2.56 N
1032.jpg	10.32 N

The .jpg file extension should be removed before converting the filename into a force value.

Sample Definition

One sample is defined as one complete tactile interaction.

Each sample contains:

1. three tactile deformation images acquired during the interaction;
2. the z-axis contact force corresponding to each image;
3. one geometric shape label;
4. one Shore A hardness label;
5. one interaction index.

The three images stored in the same interaction directory are temporally related observations from the same interaction and should not be treated as three independent interaction samples.

The dataset therefore contains 901 interaction samples and 2,703 tactile deformation images.

Shape Categories

HUST-SOFT contains three geometric shape categories: angle, circular, and edge.

Shape category	Shore A hardness values	Interaction samples	Tactile images
angle	31, 41, 53, 64, 73, 82, 87	294	882
circular	27, 39, 49, 62, 71, 79	264	792
edge	31, 41, 53, 64, 73, 82, 87	343	1,029
Total	13 distinct hardness levels	901	2,703

Angle

The angle category contains tactile interactions with angle-shaped geometric features.

• Number of interaction samples: 294;
• Number of tactile images: 882;
• Available Shore A hardness values: 31, 41, 53, 64, 73, 82, and 87.

Circular

The circular category contains tactile interactions with circular geometric features.

• Number of interaction samples: 264;
• Number of tactile images: 792;
• Available Shore A hardness values: 27, 39, 49, 62, 71, and 79.

Edge

The edge category contains tactile interactions with edge-shaped geometric features.

• Number of interaction samples: 343;
• Number of tactile images: 1,029;
• Available Shore A hardness values: 31, 41, 53, 64, 73, 82, and 87.

Compliance Categories

Material compliance is represented by the Shore A hardness of each silicone rubber specimen.

The specimens were fabricated from silicone rubber polymers with varying cross-linking ratios to finely tune their inherent stiffness. Lower Shore A hardness values generally indicate softer and more compliant specimens, whereas higher Shore A hardness values indicate stiffer specimens.

The dataset contains the following 13 Shore A hardness levels:

27, 31, 39, 41, 49, 53, 62, 64, 71, 73, 79, 82, and 87

The distribution of interaction samples across the compliance categories is shown below.

Shore A hardness	Available shapes	Interaction samples	Tactile images
27	circular	44	132
31	angle, edge	85	255
39	circular	44	132
41	angle, edge	92	276
49	circular	44	132
53	angle, edge	92	276
62	circular	44	132
64	angle, edge	92	276
71	circular	44	132
73	angle, edge	92	276
79	circular	44	132
82	angle, edge	92	276
87	angle, edge	92	276
Total	—	901	2,703

The Shore A hardness values of 27, 39, 49, 62, 71, and 79 occur only in the circular category. Each of these hardness levels contains 44 interaction samples.

Shore A 31 occurs in the angle and edge categories and contains 85 interaction samples in total.

The Shore A hardness values of 41, 53, 64, 73, 82, and 87 also occur in the angle and edge categories. Each of these hardness levels contains 92 interaction samples in total.

Valid Shape–Compliance Combinations

The valid shape and Shore A hardness combinations are:

angle:
31, 41, 53, 64, 73, 82, 87

circular:
27, 39, 49, 62, 71, 79

edge:
31, 41, 53, 64, 73, 82, 87

The dataset does not contain angle or edge samples with Shore A hardness values of 27, 39, 49, 62, 71, or 79.

The dataset does not contain circular samples with Shore A hardness values of 31, 41, 53, 64, 73, 82, or 87.

Data Fields

The labels and measurements associated with each image can be obtained from its relative file path.

Field	Description	Example
shape	Geometric shape category	angle
shore_a_hardness	Shore A hardness of the silicone rubber specimen	31
interaction_index	Identifier of an individual tactile interaction	1
image	Tactile deformation image	256.jpg
contact_force_z	Z-axis contact force in newtons	2.56

For the following file:

HUST-SOFT/angle/31/1/256.jpg

the corresponding data fields are:

shape: angle
shore_a_hardness: 31
interaction_index: 1
contact_force_z: 2.56 N

Loading the Dataset

The following Python example recursively reads the JPEG images and extracts the associated labels and contact-force values from the file paths:

from pathlib import Path

dataset_root = Path("HUST-SOFT")
records = []

for image_path in dataset_root.rglob("*.jpg"):
    relative_path = image_path.relative_to(dataset_root)

    if len(relative_path.parts) != 4:
        continue

    shape, hardness, interaction_index, filename = relative_path.parts

    record = {
        "image_path": str(image_path),
        "shape": shape,
        "shore_a_hardness": int(hardness),
        "interaction_index": int(interaction_index),
        "contact_force_z": float(Path(filename).stem) / 100.0,
    }

    records.append(record)

print(f"Number of tactile images: {len(records)}")
print(records[0])

An individual tactile deformation image can be loaded using Pillow:

from PIL import Image

image = Image.open(records[0]["image_path"]).convert("RGB")

print(image.size)
# (640, 480)

The three frames belonging to the same interaction can be grouped using the shape category, Shore A hardness, and interaction index:

from collections import defaultdict

interaction_samples = defaultdict(list)

for record in records:
    interaction_key = (
        record["shape"],
        record["shore_a_hardness"],
        record["interaction_index"],
    )

    interaction_samples[interaction_key].append(record)

print(f"Number of interaction samples: {len(interaction_samples)}")

The three images within each interaction should be ordered according to the original acquisition sequence when temporal modelling is required. The contact-force value encoded in the filename should not automatically be interpreted as a frame index.

Intended Uses

HUST-SOFT is intended to support research in:

• vision-based tactile sensing;
• material compliance perception;
• Shore hardness recognition;
• tactile contact-force estimation;
• tactile deformation analysis;
• geometric-feature recognition;
• temporal tactile representation learning;
• joint perception of geometry, compliance, and contact force;
• robotic interaction with compliant objects;
• multimodal perception;
• embodied intelligence.

The dataset can support classification, regression, representation learning, sequence modelling, and multimodal learning tasks.

Possible Research Tasks

Shape Classification

Predict one of the following shape categories from tactile deformation images:

angle, circular, edge

Because the circular category contains a different set of Shore A hardness values from the angle and edge categories, researchers should control for compliance-related information when evaluating shape-classification performance.

Compliance Classification

Predict the Shore A hardness category of a specimen from a tactile deformation image or a three-frame tactile sequence.

The complete dataset contains 13 Shore A hardness categories. However, the hardness labels are associated with restricted shape categories.

For shape-independent compliance classification, the angle and edge categories are more suitable because they share the same seven Shore A hardness values:

31, 41, 53, 64, 73, 82, and 87

Compliance Regression

Estimate Shore A hardness or another continuous compliance representation from tactile deformation images and contact-force measurements.

Researchers should consider whether a model is learning compliance-related deformation characteristics or exploiting correlations between the shape and hardness labels.

Contact-Force Estimation

Estimate the z-axis contact force from a tactile deformation image.

The contact-force target is obtained by dividing the numerical image filename by 100.

Joint Shape and Compliance Recognition

Simultaneously predict the geometric shape and Shore A hardness of the contacted specimen.

Evaluation protocols should explicitly state whether the tested shape–hardness combinations were observed during training.

Temporal Tactile Modelling

Use the three tactile deformation images and their corresponding z-axis contact forces to model the interaction process.

The three frames from the same interaction should be processed as one temporally related sample.

Ethical Considerations

The dataset contains tactile measurements of laboratory-fabricated silicone rubber specimens.

It does not contain personal information, human-subject data, facial images, biometric identifiers, or other personally identifiable information.

Users are responsible for ensuring that downstream applications comply with applicable ethical, legal, and safety requirements.

Citation

If you use HUST-SOFT in your research, please cite the associated paper:

@article{hust_soft,
  title   = {[Insert the title of the associated paper]},
  author  = {[Insert the author list]},
  journal = {[Insert the journal name]},
  year    = {[Insert the publication year]},
  doi     = {[Insert the DOI after publication]}
}

The citation information will be updated after the associated paper is formally published.

Dataset Authors

HUST-SOFT was developed by researchers at Huazhong University of Science and Technology.