Low-Light Object Detection via Image-Level Illumination-Conditioned Modulation

Abstract

Object detection in low-light scenes is challenged by weak contrast, amplified noise, and spatially nonuniform exposure, which jointly degrade feature representation and increase missed detections. Existing enhancement-first pipelines can introduce additional computation or alter task-relevant appearance, while detectors with fixed feature transformations cannot explicitly adapt to the illumination state of each input. This article proposes Illumination-Conditioned Modulation YOLO (ICM-YOLO), a compact detector that incorporates illumination awareness at both the image and prediction levels. An Image-Level Illumination Descriptor provides a shared global condition, Conditioned Dynamic Modulation C3k2 adapts semantically enriched neck features, and a Classification-Oriented Luminance-Ratio Adaptation Head recalibrates category evidence under local exposure variation while preserving the localization pathway. Experiments on ExDark show that ICM-YOLO achieves 63.75% mAP@50, improving the YOLO26n baseline by 1.32 percentage points, with an approximately 2.0% increase in parameters. Ablation results support the effectiveness of classification-oriented adaptation and selective neck conditioning. On visible-only M3FD, ICM-YOLO yields a small gain over YOLO26n. These results indicate that selective illumination conditioning offers an effective and lightweight approach to low-light object detection.