Street sweeper: detecting and removing cars in street view images

Abstract

To protect privacy of individuals or companies that may be leaked in street view images, we present a system to automatically detect and remove cars as if they had never been there. Although street view service providers have made efforts on blurring human faces and license plates, we argue that remaining features, such as license numbers and phone numbers printed on car bodies, could cause privacy leak. Given a sequence of street view images, this system first detects cars by the deformable part model, and then determines foreground/background seeds for the GrabCut image segmentation module in order to facilitate automatic car segmentation. After removing cars, an exemplar-based inpainting method is developed with special designs on filling priority determination and road structure propagation. Hierarchical texture propagation and randomized texture propagation are integrated to implement the inpainting process, so that aesthetically pleasing inpainting results as well as privacy protection can be accomplished.