TWIST: trains under weather, illumination, and seasonal time
Publikationstyp
Conference Paper
Date Issued
2026-06
Sprache
English
Author(s)
Stenger, Andre
Arkenberg, Til
Start Page
719
End Page
725
Citation
22nd Annual International Conference on Distributed Computing in Smart Systems and the Internet of Things, DCOSS-IoT 2026
Publisher DOI
Scopus ID
Publisher
IEEE
ISBN of container
979-8-3315-4670-0
979-8-3315-4671-7
Rail transport is a cornerstone of sustainable and climate-friendly mobility. With a shift towards autonomous transport not only on the road but also on rails, AI-based track monitoring systems play an important role both for operating trains as well as for maintenance along the track. To build respective applications, we require high-quality datasets capturing real-world environmental variability not currently covered by any existing dataset containing images of trains. To address this gap, we create and publicly release TWIST, a domain-specific train image dataset collected across different seasons, comprising 38,000 real-world images at 640 × 480 resolution. TWIST captures trains under a wide range of weather and illumination conditions, including rain, snow, fog, low light, and night scenes, which are largely absent from existing public datasets and critically affect model robustness in deployment scenarios.To demonstrate the practical value of TWIST, we train a compact YOLOv8s object detection model from scratch and compare it against a pre-trained baseline. The model achieves a mean Average Precision (mAP@.50) of 60.4%, comparable to the baseline value of 60.8%, while substantially improving average precision for train detection (AP@.50) from 75.9% to 89.1%. These results highlight the importance of realistic data collection for robust railway monitoring.
Subjects
dataset
object detection
railway monitoring
YOLO
DDC Class
625: Road and Railroad
004: Computer Sciences