Person Detection in Low-Light Environments: Evaluation of an Annotation-Free Approach

Abstract

This work investigates how to modify an existing dataset to enhance the capability of a predictive model' to detect objects in low-light environments. From an initial set of images captured under daylight conditions, synthetic darkened versions were generated. Subsequently, various training scenarios were organized to evaluate the impact of image transformation techniques and data augmentation strategies on predictive performance. The experiments performed demonstrate that the combination of original images and artificially darkened versions, coupled with data augmentation and neutral background images, results in more stable and accurate models for low-light detection, with negligible performance degradation in visible-light scenarios. The results highlight that it is possible to improve detector performance in real-world low-light settings without the need for new image acquisitions or additional annotation.