ADE20K SceneParsing

The 150-class ADE20K SceneParsing benchmark (the ADEChallengeData2016 release), packed as parquet so that a project can load it without unpacking the zip and wiring up paths first.

from datasets import load_dataset

ade = load_dataset("shijli/ade20k", "segmentation")   # 20210 train / 2000 val

Columns

column
id	the release's image id (ADE_train_00000001), so any row joins back to a local copy
image	the JPEG, byte for byte as released
mask	the annotation PNG, byte for byte as released
scene	the scene category from sceneCategories.txt (airport_terminal)

The challenge's test images have no public labels, so there is no test split.

Masks are labels, not pictures

A mask pixel's value is the class id. Masks are single-channel (mode L) PNGs, exactly as the release ships them: 0 is "other" (pixels outside the 150 classes) and 1–150 are the classes of objectInfo150.txt, in its order (1 wall, 2 building, 3 sky, ..., 150 flag).

import numpy as np
mask = np.array(ade["validation"][0]["mask"])   # uint8, values in {0..150}

Whether 0 is a class or ignored is the evaluation protocol's decision, so it is left as is. The common protocol ignores it and scores the 150 classes: map 0 to 255 and subtract 1 from the rest (the reduce_zero_label of mmsegmentation).

How this was packed

create_dataset.py in this repo is the script that produced these parquet files, from an unpacked ADEChallengeData2016 folder. Images and masks are stored as the original files' bytes. After packing, every one of the 22210 masks was decoded and compared pixel for pixel against its source file, and every image's bytes against its source file.

The official Hub copy, zhoubolei/scene_parse_150, is a loading script that recent versions of datasets no longer run, and its auto-converted parquet does not keep the image ids.

License and attribution

The images come from the SUN and Places databases and remain under the copyright of their respective owners. They are distributed for non-commercial research and educational use under the terms of the ADE20K dataset. This packaging adds no rights: use it under those same terms and cite the original work.

@inproceedings{zhou2017scene,
    title={Scene Parsing through ADE20K Dataset},
    author={Zhou, Bolei and Zhao, Hang and Puig, Xavier and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2017}
}
@article{zhou2019semantic,
    title={Semantic Understanding of Scenes through the ADE20K Dataset},
    author={Zhou, Bolei and Zhao, Hang and Puig, Xavier and Xiao, Tete and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
    journal={International Journal of Computer Vision},
    volume={127}, number={3}, pages={302--321}, year={2019}
}