README.md

---
license: cc-by-nc-4.0
pretty_name: ThinObject-5K
task_categories:
- image-segmentation
tags:
- thin-object-segmentation
- saliency
- matting
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

# ThinObject-5K

ThinObject-5K is a high-resolution dataset for **thin object segmentation**, containing 5,748 images with pixel-accurate binary ground-truth masks that emphasize thin structures (e.g. wires, legs, handles, wineglass stems, antennae).

## Splits

| Split | Rows | Source list |
|-------|------|-------------|
| train | 5248 | official `trainval.txt` |
| test  | 500  | official `test.txt` |

The two splits are disjoint and together cover all 5,748 image/mask pairs. The original repository additionally provides a `train.txt` (4,748) / validation (500) partition; the validation subset is folded into the `train` split here and is recoverable from the original lists if needed.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `image` | `Image` | RGB photograph (JPEG) |
| `mask`  | `Image` | Single-channel (mode L) binary ground-truth segmentation mask, same resolution as the image |

## Source & Credit

This dataset was introduced in:

> **Deep Interactive Thin Object Selection**
> Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, Jiashi Feng. WACV 2021.

Original repository: https://github.com/liewjunhao/thin-object-selection

Original data (Google Drive) is redistributed here for convenience. All credit belongs to the original authors. Please cite the paper above when using this dataset.

## License

Released under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license, matching the license of the original `thin-object-selection` repository. Non-commercial use only.