Nighttime Object Detection with Denoising Diffusion-Probabilistic Models

Object detection is essential for road safety, aiding drivers in identifying vehicles, pedestrians, and other road objects. However, nighttime detection remains challenging due to low visibility impacting the accuracy of current object detection models. This paper proposes a novel approach that uses a denoising diffusion-probabilistic model to enhance nighttime object detection performance. It is trained for conditional image translation, converting nighttime images into daytime images through a forward process that adds Gaussian noise. The reverse process predicts and removes the added noise to reconstruct the daytime image. Experimental results indicate that this method significantly improves vehicle detection accuracy at night compared to state-of-the-art detectors.