Part 2 of a synthetic dataset containing 175,428 RGB images of humans generated using Unity's Perception package. It includes semantic segmentation counterparts and JSON annotations for 2D/3D bounding boxes and COCO-format keypoints, created by CERN and the European Commission. The data was generated across 8 environments, 33 human models, 4 lighting conditions, and varied camera distances and angles.
Use Cases
- Train semantic segmentation models based on the provided segmentation image counterparts.
- Develop human pose estimation algorithms using the COCO-format keypoint annotations.
- Benchmark object detection performance using the provided 2D and 3D bounding box ground truth.
- Test model robustness across varied lighting conditions and camera angles described in the dataset.
Strengths
- Large scale with 175,428 RGB images and corresponding annotations.
- Multi-format annotations include semantic segmentation, 2D/3D bounding boxes, and keypoints.
- Systematic variation across 8 environments, 33 humans, 4 lighting conditions, and numerous camera positions.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- The dataset does not contain images for every possible combination of camera distances and angles due to collisions.
Provenance
- Source
- CERN - European Organization for Nuclear Research and European Commission
- Collection Method
- Generated synthetically using Unity's Perception package.
- Freshness
- Last updated 2026-04-23 00:00:00; freshness should be verified.