AMOS: Archive of Many Outdoor Scenes

Overview

The Archive of Many Outdoor Scenes (AMOS) dataset contains more than 20 million images taken from over 500 webcameras located around the world, the vast majority in the United States. Construction of AMOS began in March, 2006 and continues to this day. This dataset is unique in that it contains images from significantly more scenes than in previous work.

The cameras in the dataset were selected by a group of graduate and undergraduate students using a standard web search engine. Images from each camera are captured several times per hour using a custom web crawler that ignores duplicate images and records the capture time. The images from all cameras are 24-bit JPEG files that vary in size from 316 x 240 to 2048 x 1536, with the majority being 320 x 240.

Locations of the cameras in AMOS that are located in the continental United States.

A montage image created from all the images from a single camera. Rows of images correspond to the time of year and columns of images correspond to the time of day.

A montage of false-color images created using principal component analysis.

Why collect so many images?

One impetus for collecting these images was to provide a dataset to enable empirical assessment of our ideas regarding the statistics of natural scenes captured from static cameras. These statistics have not been well studied and are of great importance to surveillance algorithm and application development.

Scene Montage Images

We created a large number (currently 170) of montage images that summarize a year of images of a scene. More are available via the AMOS flickr collection.

www.flickr.com

Dataset Access

Several subsets of AMOS are available for experimenation:

All the images captured during a one week period in June 2006 for four cameras: 4, 15, 82, and 190.
One image captured around noon for many days for cameras 4, 15, 82, and 190: noon images.

The dataset details page contains information about obtaining the complete dataset.

Publications

Consistent Temporal Variations in Many Outdoor Scenes (wmv, mpg) [Please cite this paper if you use the AMOS dataset in your work.]
Geolocating Static Cameras: Determining the geographic location of a camera using cues based on the motion of the sun and local weather conditions. See the project website for more information.
Toward fully automatic geo-location and geo-orientation of static outdoor cameras: Initial work on determining the camera orientation using the clear sky as a calibration target.
Adventures in archiving and using three years of webcam images: New visualizations and extracting signals from webcam time-lapses.
The global network of outdoor webcams: properties and applications: Properties of 18,000 webcams and an application to phenology.

People

The following people have contributed to the collection, presentation, and analysis of the AMOS dataset.

Nathan Jacobs
Robert Pless
Scott Satkin
Nathaniel Roman
Richard Speyer

Related Work

Geolocating Static Cameras : localize cameras without matching image features and only using diurnal and weather related variations.
Columbia University : WILD: Weather and Illumination Database : High quality images of a single urban scene taken every hour for a year.
Spatially-localized Moving Object Shape Priors : learn the typical shapes of objects at every location in the image.

Acknowledgements

This project is supported under NSF IIS 0546383: "CAREER: Passive Vision, What Can Be Learned by a Stationary Observer". Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

This project also gratefully acknowledges AWS Convergence Technologies Inc. for allowing us to archive a collection of images from the weatherbug camera network.