TransFrag27K: Transparent Fragments Dataset

Paper

Transparent Fragments Contour Estimation via Visual-Tactile Fusion for Autonomous Reassembly

Code

Dataset Summary

TransFrag27K is the first large-scale transparent fragments dataset, which contains 27,000 images and masks at a resolution of 640×480. The dataset covers fragments of common everyday glassware and incorporates more than 150 background textures and 100 HDRI environment lightings.

Automated Synthetic Dataset Generation

Transparent objects, being a special category, have refractive and transmissive material properties that make their visual features highly sensitive to environmental lighting and background. In real-world scenarios, collecting data of transparent fragments is challenging because they are difficult to perceive and exhibit highly diverse and random shapes, which makes accurate annotation prone to errors.

To address this, we designed an automated dataset generation pipeline in Blender:

• Objects are randomly fractured using the Cell Fracture add-on.
• Parametric scripts batch-adjust lighting, backgrounds, and camera poses.
• Rendering is performed automatically to output paired RGB images and binary masks.

The Blender pipeline used to generate TransFrag27K is generalizable to arbitrary objects placed on a horizontal plane (e.g., tabletop or ground scenarios), with randomized camera viewpoints and adjustable environmental configurations. For implementation details, please refer to our GitHub repository.

---

Supported Tasks

• Semantic Segmentation for various transparent fragments.

---

Dataset Structure

In our released dataset, to facilitate subsequent customized processing, we organize each category’s data in the following structure:

├─TransFrag27K
│ ├─Planar1
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Planar2
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Curved1
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Curved2
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Irregular1
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Irregular2
│ │ ├─anno_mask
│ │ └─rgb
│ ├─Irregular3
│ │ ├─anno_mask
│ │ └─rgb

We mainly organize the dataset according to the shape classes of transparent fragments:

• Planar
  Mainly includes fragments from flat regions such as dish bottoms and glass bases.
• Curved
  Mainly includes fragments from objects with cylindrical or spherical curvature, such as cups, bottles, and bowls.
• Irregular
  Mainly includes fragments with multiple curvature patterns or discontinuous surfaces, such as the intersection of a cup wall and bottom, special bottle necks, wine glass stems, and handles.

---

Citation

If you find this dataset or the associated work useful for your research, please cite the paper:

@misc{lin2026transparentfragmentscontourestimation,
      title={Transparent Fragments Contour Estimation via Visual-Tactile Fusion for Autonomous Reassembly}, 
      author={Qihao Lin and Borui Chen and Yuping Zhou and Jianing Wu and Yulan Guo and Weishi Zheng and Chongkun Xia},
      year={2026},
      eprint={2603.20290},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.20290}, 
}