ADE20K

ADE20K is a large-scale image dataset used for semantic segmentation and scene parsing. It provides pixel-accurate annotations for a diverse collection of everyday scenes, enabling detailed labeling of both objects and region categories. The dataset is widely used as a benchmark to develop and evaluate scene understanding algorithms.

Size and labeling: ADE20K contains about 20,000 images annotated with labels for 150 semantic categories. The label set includes a mix of “things” (distinct objects) and “stuff” (amorphous regions such as sky or grass). Annotations are at the pixel level, allowing precise segmentation of each image and facilitating fine-grained evaluation across many classes.

Origin and splits: The dataset was introduced in 2017 by researchers to provide a challenging, diverse benchmark for scene parsing. It is distributed with training and validation splits; a held-out test set is used for external evaluation, with performance typically reported as mean intersection-over-union (mIoU) across categories.

Impact and usage: ADE20K serves as a standard benchmark for state-of-the-art semantic segmentation models and related tasks. It has driven advances in architectures and training strategies to cope with large label sets and complex, cluttered scenes, and is commonly used for transfer learning and methodological comparisons in computer vision research.