Dataset Curation for a Domain-specific People Detection System
Benkert J, Wagner P, Zinnen M, Christlein V, Meisen T (2026)
Publication Type: Conference contribution
Publication year: 2026
Journal
Publisher: SPIE
Book Volume: 14114
Conference Proceedings Title: Proceedings of SPIE - The International Society for Optical Engineering
Event location: Paris, FRA
ISBN: 9798902321873
DOI: 10.1117/12.3089052
Abstract
Container terminals are moving toward full automation, yet tasks such as truck maneuvering, container unlocking, and twist‑lock removal still necessitate human involvement, creating safety hazards at the interface between workers and automated cranes. Although SOTA people‑detection models can process multiple camera feeds in real time, their performance degrades sharply when deployed in port environments due to a pronounced domain mismatch with standard training datasets. Traditionally, this domain gap is bridged by manually annotating data from the target domain and finetuning the model accordingly. In port environments, which are considered critical infrastructure, capturing and annotating new images for supervised training is often prohibited due to data privacy concerns. Therefore, this work investigates the extent to which a state-of-the-art, lightweight object detection model (YOLOv8) can automatically adapt to the new target environment without requiring labeled images from that environment. To achieve this goal, a curated dataset is created which is used to fine-tune the object detection model. The curated dataset is composed of images from public datasets that closely resemble the desired target domain. Since only the feature maps are needed to create this curated dataset, the data privacy requirements can be adhered to. As we show, by using our dataset curation and model fine-tuning process improved model performance can be achieved with minimal training effort. Furthermore, data privacy and security are preserved, as no labeled images from the target domain are required.
Authors with CRIS profile
Pascal Wagner
Chair of Visual Computing (LGDV)
Mathias Zinnen
Lehrstuhl für Informatik 5 (Mustererkennung)
Vincent Christlein
Lehrstuhl für Informatik 5 (Mustererkennung)
Involved external institutions
Germany (DE)How to cite
APA:
Benkert, J., Wagner, P., Zinnen, M., Christlein, V., & Meisen, T. (2026). Dataset Curation for a Domain-specific People Detection System. In Wolfgang Osten (Eds.), Proceedings of SPIE - The International Society for Optical Engineering. Paris, FRA: SPIE.
MLA:
Benkert, Johannes, et al. "Dataset Curation for a Domain-specific People Detection System." Proceedings of the 18th International Conference on Machine Vision, ICMV 2025, Paris, FRA Ed. Wolfgang Osten, SPIE, 2026.
BibTeX: Download