Change Detection from a Street Image Pair using CNN Features and Superpixel Segmentation

This paper proposes a method for detecting changes of a scene using a pair of its vehicular, omnidirectional images. Previous approaches to the problem require the use of a 3D scene model and/or pixel-level registration between different time images. They are also computationally costly for estimating city-scale changes. We propose a novel change detection method that uses features of convolutional neural network (CNN) in combination with superpixel segmentation. Comparison of CNN features gives a lowresolution map of scene changes that is robust to illumination changes and viewpoint differences. Superpixel segmentation of the scene images is integrated with this lowresolution map to estimate precise segmentation boundaries of the changes. Our motivation is to develop a method for detecting city-scale changes, which can be used for visualization of damages of a natural disaster and subsequent recovery processes as well as for the purpose of maintaining/updating the 3D model of a city. We have created a dataset named Panoramic Change Detection Dataset, which will be made publicly available for evaluating the performances of change detection methods in these scenarios. The experimental results using the dataset show the effectiveness of our approach.