Investigation on Different Color Spaces on Faster RCNN for Night-Time Human Occupancy Modelling

Convolutional Neural Network (CNN) has been shown to have worse performance for images with low illumination. This means that object detection methods using CNN, such as Faster RCNN, will have problem detecting humans during low lighting conditions. This will lower the accuracy for human occupancy modelling during night-time. In general, most methods that has been employed to improve the performance of Faster RCNN for night-time human detection depends on RGB color space. Currently, there are few papers investigating the effect of using various color spaces including RGB color space to perform Faster RCNN night-time human detection. The objective this paper is to investigate the effect of using various color space on the performance of Faster RCNN night-time human detection. This paper will be investigating the effect of using RGB, CIEXYZ, and CIELAB color space on night-time human detection. KAIST multispectral pedestrian dataset will be used for training and evaluation. Faster RCNN will use Resnet101 base network and using ImageNet pretrained weight trained using RGB images. Using log-average miss rate as evaluation metric, night-time human detection using CIEXYZ achieves 54.93% while RGB achieves 54.42%. The results of this paper show that CIEXYZ have similar performance to RGB for night-time human detection.