Skip to Main Content
In this work, we present a technique for efficient and robust estimation of the exact location and orientation of a photo capture device in a large data set. The provided data set includes a set of photos and the associated information from GPS and orientation sensor. This attached metadata is noisy and lacks precision. Our strategy to correct this uncertain data is based on the data fusion between measurement model, derived from sensor data, and signal model given by the computer vision algorithms. Based on the retrieved information from multiple views of a scene we make a grid of images. Our robust feature detection and matching between images result in finding a reliable transformation. Consequently, relative location and orientation of the data set construct the signal model. On the other hand, information extracted from the single images combined with the measurement data make the measurement model. Finally, Kalman filter is used to fuse these two models iteratively and enhance the estimation of the ground truth(GT) location and orientation. Practically, this approach can help us to design a photo browsing system from a huge collection of photos, enabling 3D navigation and exploration of our huge data set.