Datasets:

nobg
/

figaro1k

task_categories:
  - image-segmentation
task_ids:
  - semantic-segmentation
tags:
  - hair-segmentation
  - matting
  - background-removal
size_categories:
  - 1K<n<10K
license: mit

Figaro1k

1,050 unconstrained photographs with pixel-level hair segmentation masks, spanning seven hairstyle classes. Prepared for nobg.

Usage

from datasets import load_dataset

ds = load_dataset("nobg/figaro1k")
example = ds["train"][0]
example["image"]                      # PIL RGB photograph
example["mask"]                       # PIL L mask, 0 = background, 255 = hair
ds["train"].features["hairstyle"].int2str(example["hairstyle"])  # e.g. "straight"

Splits

Split	Examples	Per hairstyle
`train`	840	120
`test`	210	30

The split is the original Training/Testing partition shipped with the dataset, not a re-split. Verified: no frame index appears in both splits, and all 1,050 frames are present.

Fields

Field	Type	Notes
`image`	`Image`	RGB photograph, original resolution — sizes vary, do not assume a fixed shape
`mask`	`Image`	Single-channel, `0` = background / `255` = hair, same dimensions as `image`
`frame`	`int32`	Original frame index (1–1050), from the source filename
`hairstyle`	`ClassLabel`	`straight`, `wavy`, `curly`, `kinky`, `braids`, `dreadlocks`, `short-men`
`width`, `height`	`int32`	Convenience copies of the image dimensions

Preparation notes

Built from the upstream Figaro1k.zip (Original/*.jpg + GT/*.pbm). Three things are worth knowing if you compare against other conversions:

Pairing is by filename (FrameNNNNN-org.jpg ↔ FrameNNNNN-gt.pbm), not by sorted directory position. The commonly-referenced loader in YBIGTA/pytorch-hair-segmentation pairs positionally, and some copies of the archive ship duplicate (1).pbm masks that silently misalign every later pair.
Masks are 255 = hair. The source is 1-bit PBM, whose polarity is easy to invert by accident. Confirmed here two ways: masks cover only ~2 % of the image border, and pixels inside the mask are markedly darker than outside (mean luminance 98 vs 154).
No resizing, cropping or normalization was applied; images are stored at original resolution. Hair covers ~34–41 % of pixels on average because the source images are tightly cropped around heads.

hairstyle is recovered from the frame index in blocks of 150 (frames 1–150 straight, 151–300 wavy, …), per the class ranges documented upstream. The resulting per-class counts come out exactly balanced, which cross-checks the mapping.

Citation

Figaro1k is released by the original authors for research purposes; see the project page, we also acquired the data thanks to the work from YBIGTA. This repository redistributes the images in a converted format and claims no additional rights over them. Cite the original work:

@inproceedings{svanera2016figaro,
  title={Figaro, hair detection and segmentation in the wild},
  author={Svanera, Michele and Muhammad, Umar Riaz and Leonardi, Riccardo and Benini, Sergio},
  booktitle={2016 IEEE International Conference on Image Processing (ICIP)},
  pages={933--937},
  year={2016},
  organization={IEEE}
}