Camera geolocation from mountain images

This paper presents a novel approach to camera geolocation from mountain images. Existing mountain view-based geolocation techniques use only mountain skylines and assume that such skylines are available and can be reliably extracted from query images. However, in real-life scenarios the skyline in a query image may be blurred or invisible, due to occlusions, adverse weather conditions and atmospheric effects, and poor image quality. Geolocating mountain view images with poor skylines is a challenge. In addition, when geolocating a query image, existing techniques do not estimate the camera roll angle, assuming that the roll angle is always small and its impact to the geolocation accuracy is negligible. However, in our research we have observed that even a small camera roll angle of only a few degrees can significantly alter the skyline features extracted from the query image and thus cause geolocation failure due to feature mismatching between the query and feature database. It is another challenge to reliably handle query images with non-negligible camera roll angles. In this paper, we propose a novel solution to these challenges by exploiting additional visual features extracted from mountain ridges beyond skylines and performing search-based camera roll angle estimation. Our proposed approach has been extensively tested on real-world challenging query images from five different regions in three continents. The experimental results from our proposed approach is significantly superior to those obtained from state-of-the-art skyline-based image geolocation approaches.