SLAM-based Robust Computational Rephotography

Computational Rephotography,aims to use the reference image or other information of the history to re-capture an image with the same pose as the reference image at the current moment.The objective is to capture the difference of observation scene in the time dimension.It has great application value in tasks such as history study,environmental monitoring,and change detection.It is also a important problem in computer vision.This paper focuses on the pain points of existing methods in actual work,combined with SLAM technology to further optimize and improve the rephotography method.This paper propose a fast and robust computational rephotography method to for feature failed under variable outdoor environment.First,the 3D information obtained by SLAM is used to establish the relationship between the history and the current.Then,according to the pose of the SLAM and the motion of the platform,the relationship between the hand-eye and the true scale can be estimated to accelerate the speed of rephotography.Finally,the accuracy is futher improved by combined the consistency of photometric and geometry.The experiments show that the proposed method has much higher robustness to lighting variations and more faster convergence rate.In addition,this paper propose a fast and reliable rephotography method under mobile device to address poor convenience.First,we accelerate the method by parallelization and optical flow.Then,we propose the feature subtitution and kyeframe to improve the roubst of varied illuminations.Finally,we provide the user guidance information in real time to complete the rephotography.The experiments show that the proposed method perform rephotography more convenience.Thus it can product rephotography in more tasks,and the rephotography can be better promoted.