CropCOCO Dataset

CropCOCO is a validation-only dataset of COCO val 2017 images cropped such that some keypoints annotations are outside of the image. It can be used for keypoint detection, out-of-image keypoint detection and localization, person detection and amodal person detection.

📦 Dataset Details

• Total images: 4,114
• Annotations: COCO-style (bounding boxes, human keypoints, both in and out-of-image)
• Resolution: Varies
• Format: JSON annotations + JPG images

🔍 Tasks

• Human Pose Estimation
• Person Detection
• Amodal Person Detection

📄 Citation

If you use CropCOCO in your research, please cite:

@inproceedings{probpose2025,
  title={{ProbPose: A Probabilistic Approach to 2D Human Pose Estimation}}, 
  author={Miroslav Purkrabek and Jiri Matas},
  year={2025},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
}

🧑‍💻 Authors

• Miroslav Purkrabek (personal website)
• Jiri Matas (personal website)