Abstract
Nighttime object detection poses significant challenges due to low illumination, noise, and reduced contrast, which can severely impact the performance of standard detection models. In this paper, we present NF-DETR (Night-Frequency Detection Transformer), a novel framework that leverages frequency domain information to enhance object detection in challenging nighttime environments. Our approach integrates physics-prior enhancement to improve the visibility of objects in low-light conditions, frequency domain feature extraction to capture structural information potentially lost in the spatial domain, and window cross-attention fusion that efficiently combines complementary features while reducing computational complexity, significantly improving detection performance without increasing the parameter count. Extensive experiments on two challenging nighttime detection benchmarks, BDD100K-Night and City-Night3K, demonstrate the effectiveness of our approach. Compared to strong baselines such as YOLOv8-M, YOLOv12-X, and RT-DETRv2-50, NF-DETR-L achieves improvements of up to +3.5% AP@50 and +3.7% AP@50:95 on BDD100K-Night, and +2.7% AP@50 and +1.9% AP@50:95 on City-Night3K, while maintaining competitive inference speeds. Ablation studies confirm that each proposed component contributes positively to detection performance, with their combination yielding the best results. NF-DETR offers a more robust solution for nighttime perception systems in autonomous driving and surveillance applications, effectively addressing the unique challenges of low-light object detection.