On the Challenges of Unsupervised Domain Adaptation for Object Detection in Event Streams

Articles in Press

Document Type : Original Article

Authors

ahmed fatlawi

Abedin Vahedian

Ahad Harati

Department of Computer Engineering, Faculty of Engineering, Ferdowsi University of Mashhad, Mashhad, Iran

10.22067/cke.2026.100568.1207

Abstract

Event cameras offer high temporal resolution, wide dynamic range, and robustness to motion blur, but their sparse event-based output differs fundamentally from intensity images used to train most object detectors. In addition, event-camera datasets with object-level annotations remain severely limited relative to the large labeled image datasets available for conventional vision. Together, this cross-modal mismatch and the limited availability of labeled target data make direct retraining less practical and motivate mechanisms for transferring knowledge from image-trained detectors to event data without target-domain detection labels. This study compares two practical transfer routes for a YOLOv5s detector: adapting the detector to event-derived input, the approach taken by ConfMix and SF-YOLO, or relinquishing the adaptation altogether by translating event streams into an intensity-like representation and applying the original pretrained detector. ConfMix retains labeled source supervision and adapts through confidence-guided source–target mixing, while SF-YOLO is a source-free teacher–student method driven by target-domain pseudo-labels. These adaptation strategies are evaluated against E2VID, a pretrained recurrent model that reconstructs events into intensity images. The Cityscapes dataset provides labeled conventional images for the source domain, whereas DSEC provides real automotive event-camera streams for the target domain. Overall, the results show that image-to-event transfer depends jointly on detector-facing representation, temporal support, and the adaptation mechanism. No single route is preferable under all conditions: reconstruction becomes the stronger option only when the temporal window is long enough to support it, and the ordering reverses in favour of detector adaptation at shorter windows and on an independent event dataset.

Keywords

event cameras

object detection

unsupervised domain adaptation

source-free domain adaptation

cross-modal transfer

event-to-video reconstruction

DSEC

YOLOv5

Subjects

• Machine learning

---