15,103 images across Cityscapes and PASCAL VOC benchmarks annotated with part-level segmentation masks for semantic object classes. The dataset provides a unified framework for reading, processing, and evaluating hierarchical scene labels as introduced in the CVPR 2021 paper.
Use Cases
- Train hierarchical segmentation models using the part-level mask annotations
- Evaluate model performance on scene parsing using the provided CVPR 2021 evaluation tools
- Visualize complex object-part hierarchies using the included Python visualization scripts
Strengths
- Includes 5,000 images from Cityscapes and 10,103 images from PASCAL VOC
- Provides a unified data format for semantic, instance, and part-level segmentation masks
- Contains official evaluation scripts for the Panoptic Quality (PQ) metric extended to parts