Efficient and Distortion-Aware Fisheye Object Detection for Edge Devices

Bao Tran Gia, Tuong Bui Cong Khanh, Tam Le Thi Thanh, Hien Ho Trong, Tien Do, Thanh Duc Ngo, Duy-Dinh Le, Shin'ichi Satoh; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 5469-5475

Fisheye cameras are increasingly employed in traffic surveillance systems due to their ultra-wide field of view, which allows for extensive scene coverage with a reduced number of devices. However, object detection in fisheye images remains challenging for real-time applications. Many recent methods rely on multi-model ensembles to boost accuracy, but these approaches incur high computational costs and latency. Moreover, fisheye distortion causes peripheral objects to appear small and warped, making them difficult to detect and leading to unstable model convergence. In this work, we present an efficient and scalable framework for fisheye object detection that addresses these challenges from both algorithmic and system perspectives. We propose a distortion-aware learning strategy that improves detection performance in highly distorted regions and a spatially guided augmentation method to enhance robustness. To support real-time deployment, we implement a set of GPU-level runtime optimizations that significantly reduce memory usage and latency. Extensive experiments on benchmark datasets demonstrate that our method achieves competitive accuracy while maintaining high inference speed (FPS), making it well-suited for edge-device deployment in practical traffic surveillance systems.