22,000 annotated scene-centric images partitioned into 20,000 training and 2,000 validation samples across 150 semantic categories. The data includes exhaustive annotations for both background elements like sky and road, as well as discrete objects like persons and beds.
Use Cases
- Train deep learning models for semantic segmentation using the 150 provided category labels.
- Benchmark scene parsing algorithm performance using the standardized 2,000-image validation split.
- Develop hierarchical segmentation models that leverage the exhaustive annotations of objects and their constituent parts.
Strengths
- 20,000 training and 2,000 validation images derived from the ADE20K Dataset.
- 150 semantic categories including 'stuff' (sky, road, grass) and discrete objects (person, car, bed).
- Exhaustive annotations covering both whole objects and specific object parts.
- Non-uniform object distribution that mimics natural daily scene occurrences.