SLOOP is a pattern retrieval engine for Animal Biometrics that uses cloud computing, machine learning and crowd sourcing to greatly improve the study of animal movement and behavior. It is also known as an Individual Identification, Photo ID, or Visual Recapture system. Sloop's primary application is in ecology for conservation and wildlife tracking, monitoring and management. Sloop is a continuous improvement from Image Retrieval and Recognition research that started in 1996 in another context. In 2002 the initial methods that formed its first version were applied to the Marbled Salamander and published in ACCV 2004. The name Sloop was coined by Chris Yang, much later, in 2008 and became the standard name for all versions after that.
Sloop was developed for two reasons. First, Sai Ravela, in 1998, became a sustainability activist, pivoting his research to be earth and environment centered (and a reason he joined EAPS). Second, it was clear that in order for scientists to get a handle on issues such as genetic variation, dispersal, diversity and movement of a species, an accurate track or capture history of individuals was needed. Historically, such counts involved a laborious process of comparing hundreds of images, often obtained by remote camera. Scientists and students spent thousands of hours poring over these images in order to identify individual animals, time arguably better spent elsewhere. SLOOP applies pattern recognition to these same images with great efficacy.
How is Sloop Different?
At the time, it wasn't clear how much of a benefit one would find. Solving the problem requires high recall and fully automatic identification, technically a far cry from reality even today, especially as new individuals enter the population pool year after year in numbers. Sloop introduced the image retrieval paradigm that provides ranked retrievals of possible matches. Initially, the user verifies the top few retriveved matches and the system learns online to get better at pulling up the rest to the top ranks. As more matches were found, it got better at finding the rest. In this way, with few verifications of the top few retrievals in each relevance feedback iterations photographs could be rapidly labeled with identifies and indexed. Sloop was the first system to show this. The approach worked even when new individuals were constantly entering the collection. A little later, as Sloop gets better over relevance feedback iterations, it is able to utilize individual-based thresholds to automate the matching process, for subseqeuent sighting/collection periods -- but this is typically treated with caution on a species-by-species basis. To obtain relevance feedback, Sloop challenges online crowds to sort and identify a small proportion of potential matches. ***
What does Sloop do well?
SLOOP learns from the citizen scientist’s skill to match images of individual animals more accurately and far faster than by previous methods. It is an early example of a system where the human uses the machine to accomplish a task more efficiently and the machine learns from sparse human input to get better at what it does.
Unlike competing systems that only work to "Re-Identify," that is, recognizes known individuals again, Sloop offers excellent "zero-shot" use in the field. It rapidly builds an indexed collection of identities and then updates them over time as new photographs are added. Sloop works in both automated and assisted modes (human-in-the-loop), determined in large measure by the errors that can be accomodated for testing an ecological or biological hypothesis.
SLOOP was originally designed for the marbled salamander of Western Massachusetts. Scientists around the globe rapidly understood its application to other species. Thus far, SLOOP has been adapted for use with Grand and Otago skinks in New Zealand and is in the process of being extended to many species.