Does Low-Light Image Enhancement Improve Object Detection? A Detection-Oriented and Cost-Aware Study on ExDark

Abstract

Low-light image enhancement (LLIE) is widely used as a preprocessing step for object detection under the assumption that visually enhanced images should improve downstream detection performance. However, whether this assumption holds for modern object detectors remains unclear. This paper presents a detection-oriented and cost-aware evaluation of representative classical and learned LLIE methods on the ExDark dataset using two off-the-shelf COCO-pretrained detectors, YOLOv8n and Faster R-CNN. In addition to raw images, we evaluate CLAHE, LIME, Zero-DCE, RetinexFormer, and a training-free hybrid enhancement pipeline. Across both detector architectures, no enhancement method consistently outperforms raw images, while a validation-tuned CLAHE baseline performs competitively with substantially more complex learned methods. Component ablation further shows that the aggressive Multi-Scale Retinex with Color Restoration (MSRCR) branch contributes most of the computational cost while providing limited detection benefit. In contrast, a lightweight conservative branch achieves detection performance comparable to raw images on YOLOv8n while reducing enhancement latency by approximately 275× compared with the full pipeline. These findings suggest that, for off-the-shelf low-light detection on ExDark, preserving detector-relevant image structure may be more important than maximizing perceptual image quality under the evaluated setting. The results further highlight the importance of evaluating low-light enhancement using downstream detection performance together with computational efficiency rather than visual quality alone.