Cataloging Birds in Their Natural Habitat

We devise a method to catalog novel objects from an image sequence even when the underlying scene exhibits sudden motion and appearance changes in consecutive frames, exemplified by the case of birds in their natural habitat. Cataloging birds in different ecosystems can provide important measures towards scientific models of global warming. However, images captured of the natural environment exhibit many visual “nuisances” that challenge standard detection and tracking methods that would allow for the cataloging of birds. We propose a method that specifically models the finescaled changes on the background due to motion, selfocclusion, and lighting changes. Regions that do not fit in this model are considered an instance of some bird. We then associate these regions with bird identities by allowing for either appearance similarity or location proximity as a guide. Birds are then clustered into visually similar groups that approximate species. Experiments show that we can maintain tracks for significantly longer periods of time as compared to classic mean shift tracking, and provide meaningful clusters for the end user.