UC Merced

The geographic and environmental interpretation of photographs is of increasing in- terest due to the growing availability of large scale datasets annotated with location information. An automatic interpretation system can assist humans in undertanding landscapes of large areas, provide cues for the scenicness of a location, or pre-filter the unrelated images for further processing, such as object and event recognition. My work mainly focuses on two problems: estimating atmospheric visibility from static images and estimating the "scenicness" of the scene depicted in an image and generating a cor- responding map. For the problem of using images to estimate atmospheric visibility, I construct the analytic mapping from image feature space, which consists of spatial contrast between pixels or frequency energy, to the optical measure of atmospheric visibility. Also, I explore a manifold learning algorithm from machine learning to obtain a more reliable estimator by constructing a graph of images based on their temporal order, since the direct mapping is very challenging. To automatically estimate the scenicness of an image, I perform experiments using global image features. A set of manually labeled images are used to learn a regressor for predicting the labels of novel images. I also investigate two extensions: first, I propose a novel composite visual-geographic location kernel which considers both the visual similarity and geographic proximaty of images; and second, I improve the accuracy of the learned regressor by incorporating unlabeled images in a graph Laplacian semi-supervised learning framework. I also investigate a novel composite visual-geographic graph Laplacian regularization term. I demonstrate the approach by using ground-level images to label how scenic a location is.