Adventures in archiving and using three years of webcam images

Recent descriptions of algorithms applied to images archived from webcams tend to underplay the challenges in working with large data sets acquired from uncontrolled webcams in real environments. In building a database of images captured from 1000 webcams, every 30 minutes for the last 3 years, we observe that these cameras have a wide variety of failure modes. This paper details steps we have taken to make this dataset more easily useful to the research community, including (a) tools for finding stable temporal segments, and stabilizing images when the camera is nearly stable, (b) visualization tools to quickly summarize a years worth of image data from one camera and to give a set of exemplars that highlight anomalies within the scene, and (c) integration with LabelMe, allowing labels of static features in one image of a scene to propagate to the thousands of other images of that scene. We also present proof-of-concept algorithms showing how this data conditioning supports several problems in inferring properties of the scene from image data.