AIM-500

Automatic Image Matting-500 — 500 high-resolution natural images with manually labelled alpha mattes, as one test split, which is what the authors published it as.

from datasets import load_dataset

ds = load_dataset("nobg/AIM-500", split="test")   # 500 rows
ds[0]["image"]       # PIL RGB, original resolution
ds[0]["mask"].convert("L")   # a real 8-bit alpha matte, not a binary mask
                             # (mode varies upstream — see the note below; convert is lossless)
ds[0]["category"]    # 'animal' | 'portrait' | 'plant' | ... (7 values)
ds[0]["type"]        # 'SO' | 'STM' | 'NS'
ds[0]["image_name"]  # 'o_004cddc9'

Why this mirror exists

The authors distribute AIM-500 as a Google Drive folder, not an archive — there is no single URL to load_dataset from, and the per-image metadata lives in a separate JSON. This mirror is that folder in parquet with the metadata joined in as two real columns, and with image and alpha bytes passed through unmodified (verified by SHA-256 on all 1000 files). The trimap/ and usr/ folders are not included: both are derived from mask and exist for trimap-based matting methods.

The masks are true soft alpha. Median 256 distinct grey levels per mask, mean 6.55 % of pixels strictly between 0 and 255 (max 93.5 %). That is the point of the set, and it is why nothing here re-encodes: a lossy or palettized round trip would destroy exactly the soft edges being measured. Downstream code should not binarize these.

One upstream property to know before you decode

The masks are not all saved in the same PIL mode: 161 are L, 293 are RGB and 46 are RGBA. Those bytes are the authors' and are passed through as-is. It matters because .convert("L") weights RGB by ITU-R 601 and discards the alpha channel — either of which would corrupt a matte.

Measured on all 339 non-L masks in this mirror: every one is grey-replicated (R == G == B at every pixel) and every RGBA one has fully opaque alpha (min == 255). So the matte lives in the colour channels, not the alpha channel, and .convert("L") reproduces it exactly — 0 differing pixels. datasets decodes these to whatever mode the file declares, so call .convert("L") (safe here) rather than assuming L, and never read the A band as the matte.

Categories and types

category	images
animal	200
portrait	100
plant	75
furniture	45
toy	36
transparent	34
fruit	10

type is the authors' difficulty axis: SO salient opaque (424), STM salient transparent/meticulous (43), NS non-salient (33). The transparent category and the STM type are the hard slices — useful for isolating where a model's alpha actually fails rather than reporting one average.

Published comparator: GenPercept reports zero-shot SAD 75.5 on AIM-500, against ViTAE-S's 112.52.

Licensing

MIT, stated by the authors in the dataset's own readme.txt ("The dataset is under MIT license") — no gate and no signed agreement. Please cite the paper.

Citation

@inproceedings{li2021deep,
  title={Deep Automatic Natural Image Matting},
  author={Li, Jizhizi and Zhang, Jing and Tao, Dacheng},
  booktitle={IJCAI},
  year={2021}
}