In the final workshop session, each session chair was charged with highlighting some of the next steps that were discussed during each session’s workshop presentations and discussions. Each session chair was asked to describe what the different disciplines (biology, computer vision, and visual analytics) should do for one another. Benjamin Richards suggested that the computer vision community provide a set of integrated software tools to extract biological data needed from existing image data; he also suggested that it develop a set of hierarchical classifiers. Mubarak Shah said that computer vision researchers should learn about the types of existing data and problems in the fisheries community. Chuck Stewart suggested that fisheries researchers identify a basic set of problems at the right scale and engage all user groups to solve it. David Jacobs emphasized the need for using real-world data, to see what can be done with imperfect tools. Hui Cheng suggested that the fisheries community prioritize a list of problems that need to be addressed.
The question was then posed to the larger audience. One participant proposed making the algorithms smart enough to understand when they do not apply well to the image under consideration; otherwise, the fisheries community may not realize when an algorithm does not match well to a given set of circumstances. Another participant emphasized the importance of dual novelty: a problem must be novel in and useful to both the fisheries community and the computer vision community for both communities to want to engage in research. Another participant suggested returning the level of uncertainty with a result, which is as important as the result
of a classifier; many current fisheries identification and classification tools do not provide a probability, only a result. Shah noted that the most common classifier, the support vector machine, provides a confidence interval.
A participant considered the potential utility of prizes and asked how these were used in the computer vision community. Computer vision does have a history of using prizes to motivate research in certain topics or on certain data sets. The cost of developing the tools to run a competition, someone noted, is far greater than the prize money itself. Another participant noted the challenge of creating the right data set in a competition; for example, if there is any bias in the data set, it can be exploited to obtain the best results, and then the results do not generalize to other data sets. Another participant cautioned that prizes do not draw a representative sampling of researchers, because established researchers may hesitate to risk their reputation in a competition.
Another participant suggested that the community focus on “low-hanging fruit”—that is, projects that would have a high likelihood of success without a large investment in resources. An example of this, the participant suggested, might be Research Experiences for Undergraduates projects.1
Several additional topics were discussed at the workshop on different occasions but not in the final concluding session. Other discussion items that were addressed by multiple speakers or participants during the course of the workshop include the following:
- Partial automation. Several speakers from the fisheries community, including Dvora Hart, Allan Hicks, Benjamin Richards, and Elizabeth Clarke, emphasized to the workshop participants that making manual measurements of fish abundance, size, and taxonomy is a laborious process that consumes human capital; the automation of such tasks would be very valuable to the fisheries community. Several participants, including Dvora Hart and Hanumant Singh, stated that even partial automation would be helpful—for example, the automatic separation of underwater images into those with fish and those with no fish, because subsequent processing need not be performed on the latter set.
- Absolute abundance measurements. Allan Hicks, Dvora Hart, and other participants from the fisheries community also stressed the importance of
1 The National Science Foundation funds research opportunities for undergraduates through its Research Experiences for Undergraduates projects. For more information, see “NSF REU Program Overview Home Page,” http://www.nsf.gov/crssprgm/reu/, accessed September 26, 2014.
tools and techniques to obtain accurate measures of absolute abundance, rather than the relative measures that are more commonly available today.
- Data sets. Several speakers and participants, including Clay Kunz and Anthony Hoogs, suggested increasing the visibility of the types of problems that currently exist in the fisheries community, because computer vision experts are unlikely to be aware of these data sets and research questions. Anthony Hoogs and Clay Kunz indicated some difficulty in the current access to fisheries data; there is no central repository for data from which computer vision and visual analytics experts can familiarize themselves with the types of data and types of problems in the fisheries community. In addition, Hoogs noted that lack of easy access to data can lead to competition, rather than collaboration, for data sets.
- Tools. Benjamin Richards and other participants stressed the need for an integrated kit of video and image processing tools that could be applied to the fisheries community. Other participants noted that many such tools are under development and are available in open source. However, maintaining such toolkits, some said, is a substantial task that would need to be formally addressed. Anthony Hoogs proposed the development of meta-learning computer vision tools to enable scientists who are not experts in computer vision or visual analytics to generate and apply different computer vision algorithms to their data.
- Engaging the community. Participants discussed novel ways to engage the user community in efforts to annotate images and track fish to add information to computer vision data training sets. Serge Belongie and Benjamin Richards both described efforts in crowdsourcing to hand annotate images, and Concetto Spampinato discussed an online game in which users identify fish in still images. In addition, concrete plans were made to hold a computer vision workshop on the topic of integrating computer vision and aquaculture operations in 2015.2
2 See the website for the First Workshop on Automated Analysis of Video Data for Wildlife Surveillance, January 9, 2015, in Waikaloa Beach, Hawaii, at http://marineresearchpartners.com/avdws2015/Home.html, accessed September 22, 2014.
This page intentionally left blank.