README.md

---
license: cc-by-nc-4.0
pretty_name: HIM-2K
task_categories:
  - image-segmentation
tags:
  - human-instance-matting
  - alpha-matte
  - matting
---

# HIM-2K

HIM-2K is a human instance matting benchmark introduced in
**"Human Instance Matting via Mutual Guidance and Multi-Instance Refinement"**
(InstMatt, Sun et al., CVPR 2022). Please cite the original authors and respect
the non-commercial (CC BY-NC 4.0) license.

## Schema

One row per image:

- `image` — the RGB photo (`datasets.Image`).
- `mask` — a **list** of per-instance alpha mattes for that image
  (`List(Image(...))`); each element is one human instance alpha, ordered by the
  instance index in the source archive.
- `subset` — `"comp"` or `"natural"` (matches the split name).

## Splits

- `comp` — 1680 composited images.
- `natural` — 320 natural images.

The `images/natural_wo_gt` portion of the original archive is **excluded** here
because it has no ground-truth alpha mattes.

## Citation

```
@inproceedings{sun2022instmatt,
  title     = {Human Instance Matting via Mutual Guidance and Multi-Instance Refinement},
  author    = {Sun, Yanan and Tang, Chi-Keung and Tai, Yu-Wing},
  booktitle = {CVPR},
  year      = {2022}
}
```