A Lightweight and Data-Centric Framework for Real-Time Object Detection in Fisheye Camera

An To Vinh, Nguyen Mai Vinh, Nguyen Luong Si, Duy Do Quoc, Duy Tran Khanh, Nguyen Nguyen Hoang, 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. 5500-5507

Fisheye cameras offer wide-area coverage in traffic surveillance but introduce severe radial distortion that complicates general object detection -- especially near image peripheries. Addressing this requires distortion-aware and semantically consistent training data. We propose a lightweight, data-centric framework for real-time and general-purpose object detection in fisheye imagery. Our method consists of three key components. First, we introduce a distortion-aware copy-paste augmentation strategy that preserves geometric alignment and contextual plausibility. Second, we construct a multi-source synthetic data pipeline combining fisheye simulation, style transfer, and background generation to enhance training diversity. Third, we design a dual-detector inference module using YOLOv11 and D-FINE, fused via Weighted Boxes Fusion for robust detection without compromising speed. The framework is model-agnostic and category-agnostic, making it easily extensible to a wide range of object types beyond traffic-specific classes. Our method currently ranks 1st on the public leaderboard of the AI City Challenge 2025 (Track 4), achieving an F1 score of 0.6493, demonstrating both accuracy and adaptability for fisheye-based applications.