Sellestia Dataset

Background-removal data for top-down food photography where the plate is part of the subject: every mask keeps the dish and its plate.

It was used to fine-tune Sellestia, a BiRefNet / Lucida fine-tune that keeps the plate where general-purpose models drop it.

Dataset Summary

• 302 annotated images, 1024×1024, split train / validation / test.
• Each sample has the photo, its alpha mask and the RGBA cutout.
• Top-down dishes on a solid orange background, on plates of many shapes, materials and colours (white, cream, grey, black, terracotta...).
• Curated by: Franck Desfrançais, Sellest
• License: CC-BY 4.0
• Intended task: image segmentation / matting (foreground = dish + plate)

Uses

Direct use

• Training, fine-tuning or benchmarking background-removal models on food photography.
• The test split is the benchmark reported in the Sellestia model card.

Out of scope

• The backgrounds are mostly solid orange: on its own, this data does not teach background removal on cluttered or real-world scenes.

Sample Usage

from datasets import load_dataset

ds = load_dataset("franckdsf/sellestia")
sample = ds["train"][0]
sample["image"]   # PIL.Image, RGB photo
sample["mask"]    # PIL.Image, L: white = keep (dish + plate), black = background
sample["cutout"]  # PIL.Image, RGBA: photo with the mask as alpha

Plain files

The same images are also in sellestia.tar, as folders ready for BiRefNet-style training (im/ + gt/):

huggingface-cli download franckdsf/sellestia sellestia.tar --repo-type dataset --local-dir .
tar -xf sellestia.tar

Dataset Structure

data/                                   Parquet, used by load_dataset and the viewer
├── train-0000{0,1}-of-00002.parquet   212 samples
├── validation-00000-of-00001.parquet   45 samples
└── test-00000-of-00001.parquet         45 samples
sellestia.tar                           the same samples as PNG files
├── train/
│   ├── im/    # RGB photos
│   ├── gt/    # masks
│   └── an/    # RGBA cutouts
├── validation/…
└── test/…

Files share the sample id as their name, e.g. train/im/<id>.png, train/gt/<id>.png.

Column	Type	Description
id	string	Sample identifier
image	image (PNG, RGB)	Original photo
mask	image (PNG, L)	Soft alpha mask, 0–255; white = dish + plate
cutout	image (PNG, RGBA)	image with mask as alpha channel

The splits are random (15% validation, 15% test) and do not overlap.

Dataset Creation

Curation rationale

General-purpose background-removal models treat the plate as background: they cut out the food alone, or leave a half-transparent ghost of the plate. Food and recipe apps usually need the whole served dish. This dataset defines that target explicitly.

Source data

• Synthetic top-down food photos generated with a text-to-image model, covering a wide range of dishes and cuisines, with the plate style varied per dish.
• No real people, places or personal data appear in the images.

Annotations

• Masks were produced with a background-removal tool, then reviewed by hand: inaccurate cutouts were corrected or removed.
• Masks keep the soft alpha at the plate's edge.

License

CC-BY 4.0: free to use, share and adapt, including commercially, with attribution.