| Internet imagery has grown in size dramatically over the last decade. These images are ubiquitous, diverse, and useful for a variety of practical vision algorithms. Unfortunately, due to their unstructured nature, these images are largely uncalibrated. In order to use these images for applications in environmental monitoring and photo-forensics, the images should ideally be calibrated, to know where an image was taken, what nonlinear transformations the camera applies to the raw sensor readings, and the underlying 3D shape of the scene. We present methods to perform this camera calibration and 3D reconstruction for outdoor scenes from a single view over time. First, we give a web-based calibration tool to calibrate for the geometric and geographic context of a camera. Second, as the sun passes over an outdoor scene, surfaces will become brighter or darker depending on the local variation of the surface with respect to the sun. Therefore, intensity-based temporal cues help to uncover the subtle 3D surface variation of a scene. Finally, cast shadows from nearby buildings and trees provide projective-distorted cues for the 3D structures of both the shadow-casting object and the object under shadow. This cast shadow analysis is extended to analyze the motion patterns of shadows, which can be used to reconstruct objects the camera never directly saw. |
