Real-Time Object Detection on Edge Devices: A Fisheye Specific DFINE

Jaewoo Park, Jihae Lee, Yunjeong Yong; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 5476-5485

In this paper, we propose Fisheye-Aware Real-Time DFINE, an end-to-end object detection system that addresses both fisheye distortion and the latency constraints of edge devices. The model consists of three key components: (i) distortion preserving augmentation, which significantly improves peripheral recognition performance. By combining radial distortion, perspective projection, and lens simulation, the detection rate of peripheral objects improved by +26.5 %p (AP50:95). (ii) Through semi-automated thresholding scheme based pseudo-labeling, the training dataset was expanded to 170,184 images, resulting in a more robust model. (iii) The distribution refinement DETR structure limited F1 degradation to only -3.68 %p even after TensorRT-based INT8 quantization, while achieving 17.1 FPS on a Jetson AGX Orin 32GB Developer Kit. In the AI City Challenge 2025 Track 4, our system achieved an F1 score of 0.6268, improving upon the baseline DFINE by +11.97 %p. Furthermore, in real-world evaluations, our system achieved real-time analysis, demonstrating both the effectiveness of single-fisheye camera-based object detection and its scalability in smart-city scenarios.