Data Augmentation with Pseudo-Infrared Night-Vision Image Conversion for Improved Nighttime Object Detection
Ryoma Ishizu, 裕教 北風, Ryo Matsumura · 2023
This paper presents a novel data augmentation method with pseudo-infrared night-vision image conversion for RGB images captured under daylight conditions. Our proposed data augmentation method is to apply gamma correction after grayscale conversion to RGB images and composites them with a mask image that simulates infrared irradiation. Our goal is to improve the accuracy of nighttime object detection using infrared night-vision cameras. We exclusively utilized data from the "person" class in the PASCAL VOC dataset and conducted a nighttime detection performance comparison between models trained on this person dataset (original dataset) and models trained on the person dataset to which we applied our proposed data augmentation method (converted dataset). The experimental results showed that the Average Precision (AP) for the model trained on the original dataset is 74.9%. In contrast, the AP for the model trained on our converted dataset is 78.7%.