KinematicEvent-HumanUpperBody 2026

Dataset Information

The KinematicEvent-HumanUpperBody (KE-HUB) dataset introduces a collection of synchronized real-world event and RGB streams. To eliminate human motion variance and accurately mimic applications where the camera itself is in motion, the data was captured utilizing a physical event sensor mounted on a cobot executing programmed trajectories. Thus ensuring precise, reproducible egomotion and a completely static subject setup. The dataset further entails facial movement in subjects, as they were asked to read short text-passages in the english language (text of each subject is provided as ground truth (gt)).

:lock: How to Access the Dataset

Due to privacy considerations, this dataset is closed-source and available upon request. To gain access, please follow these steps:

1. Download the Data Usage Agreement (DUA).
2. Fill out and sign the document.
3. Send the completed DUA via email to zml@hs-heilbronn.de & adam-theo.mueller@hs-heilbronn.de.
4. Once your request is reviewed and approved, you will receive a secure download link.

Technical Specifications

Data was captured using one IDS XCP-E event camera (f = 4 mm), mounted between two Basler ace 2 cameras (f = 8 mm, 30 FPS). The cobot used was a Universal Robots UR5e (the movement file (.urs) is provided in the dataset). The camera plane was 160 cm infront of the background, with the subjects thus in ~120 cm distance to the camera plane.

• Total Size: (extracted) ~230 GB
• Number of Subjects: 11
• Sequence Length: ~8 seconds per video (two independent videos per subject, for one to be used as gt)
• Formats Provided:
  • RGB: .mp4, .png frames, also with event in ROS 2 bag format (.mcap)
  • Events: .bin (+ metadata.json), .npy (also event debug video as .mp4)
  • Text: .txt (text-sequence read by the subject, in both videos per subject the same)
  • Robot Movement: .urs

Citation

If you use this dataset in your research, please cite our paper:

@misc{mueller2026genEventAnon,
      title={Generative Anonymization in Event Streams}, 
      author={Adam T. Müller and Mihai Kocsis and Nicolaj C. Stache},
      year={2026},
      eprint={2604.12803},
      archivePrefix={arXiv},
      doi={10.48550/arXiv.2604.12803},
      url={[https://arxiv.org/abs/2604.12803](https://arxiv.org/abs/2604.12803)}
}