Boosting Fisheye Detection with Augmentations and Ensembles

Chengzhi Qian, Jing Li, Yangyang Huang, Zhixin Lin, Yue Yao, Jianping Wang; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 5495-5499

In recent years, traffic surveillance systems have increasingly adopted fisheye cameras to achieve wide-area coverage with fewer devices. However, the severe radial distortion inherent in fisheye imagery poses significant challenges for conventional object detection algorithms. To address this, we present a robust object detection framework tailored for traffic scenarios using fisheye inputs. Our approach begins with a novel data augmentation strategy designed to better simulate the geometric characteristics of fisheye images, thereby enhancing the model's adaptability. To further enrich the training set, we incorporate pseudo labels generated by a pre-trained detector, enabling the creation of more diverse and representative training samples. On the modeling side, we integrate detectors based on the YOLOv11 family. In particular, we explore enhanced variants that incorporate CGRFPN and FDPN, which boost feature extraction and multi-scale representation. In addition to using the YOLOv11 family as our core detection architecture, we also leverage YOLOv8 in conjunction with pseudo-labeling to further improve training quality and performance. These models are then combined using robust ensemble techniques to ensure improved generalization and detection accuracy. Our proposed method achieves an F1 score of 64.05% on the 2025 AI City Challenge - Track 4, securing second place among all participating teams and demonstrating strong performance in complex real-world traffic scenarios.