Enhancing Fisheye Object Detection Using Frequency-Domain Attention and Dual Aggregation Transformer

This paper presents a comprehensive evaluation of multiple YOLO (You Only Look Once) model variants for object detection in fisheye lens images, specifically utilizing the FishEye8K dataset. We propose a novel framework that combines frequency-domain preprocessing using FFTformer and post-processing with Dual Aggregation Transformer (DAT) to enhance detection accuracy. The study benchmarks ten variants across YOLOv5, YOLOv8, and YOLOv9, evaluating their performance on five object classes: Pedestrian, Bike, Car, Truck, and Bus. Our preprocessing strategy employs FFTformer to improve image clarity and reduce noise, while the post-processing approach utilizes DAT for super-resolution enhancement. Results demonstrate significant improvements in precision, recall, mAP50, and mAP metrics across all object categories after implementing the proposed preprocessing and post-processing techniques. Notably, YOLOv5l6 exhibited superior performance in detecting smaller objects, while YOLOv8x excelled at larger object detection. The integration of FFTformer and DAT led to substantial improvements in detection accuracy, particularly for challenging scenarios involving distorted or low-quality fisheye images.