Synthetic RAW Augmentations Improve Pedestrian Detection in Low Light

A new study from arXiv demonstrates that synthetic low-light images generated via RAW augmentation can improve evaluation of pedestrian detection in autonomous driving. Researchers used a synthetic RAW image augmentation technique to create low-light samples matching a camera sensor's noise model. Performance metrics on real and synthetic low-light data were similar, indicating the AI model's robustness. The work addresses the challenge of sparse real datasets in low-density regions, enabling more continuous sampling of input space for benchmarks. The study focuses on the safety-critical case of pedestrian detection in the dark, using a state-of-the-art object detection model. The method shows potential for better characterizing model performance as a function of scene illumination.