The World Is Changing: Finding Changes on the Street

Abstract

We propose to find changes in the constantly changing world, given visual observations at street-level. In particular, we identify “long-term” changes between Google Street View images and dashcam videos captured at different months or even years. This is a challenging task, since (1) dashcam frames are not localized in world coordinate, and (2) there are many changes introduced by moving objects. We propose a robust sequence alignment method to align dashcam sequence to Street View images. Our method outperforms a strong baseline method [1] by $12\%$ mean Average Precision (AP). We also propose a novel change detection method designed to detect long-term changes. Our change detection method (${13.54}\%$) outperforms a baseline method without handling car interior and moving objects (${11.70}\%$) by $1.84\%$ (relatively $13.6\%$) in mean AP. In a controlled experiment, given manually aligned high quality Street View images, our change detection method achieves a significantly better mean AP ($45.57\%$).