Etienne Meunier
From Dataset to Algorithm, Etienne Meunier (MET’19) Teaches Machines to “Recognize” Sharks
Etienne Meunier
Freelance IT Consultant
MS in Computer Science, Concentration in Data Analytics (MET’19); MEng, ECE Graduate School of Engineering (Paris)
What compelled you to return to school and pursue a graduate education? What is your long-term objective?
I entered this program at Boston University as part of an exchange through my school in France (ECE Paris). It was a very good opportunity for me to continue to study computer science and live a new international experience. I was also really attracted by the reputation of Boston University and the content of the program. My long-term objective is to use technology and computer science to work on important and significant problems and offer what I know for valuable projects, but because there are a lot of ways to get to that point my medium-term objective is to master my domain of study (machine learning) and learn while investing in challenging projects. I am also interested in getting involved in research because I think it is a very good way to learn and focus on a topic. That is why I will start a PhD next year.
Is there a particular faculty member from your courses who has enhanced your experience at BU MET? Who and why?
Without any doubt, Dr. Eugene Pinsky enhanced my experience at Boston University a lot. He guided me in the research field and, thanks to his trust and support, I had been able to be involved in wonderful projects. He helped my friend Pierre Moreau (also a student at BU MET) and me to write our first publication on long-term power consumption forecasting. Furthermore, his course on Data Science with Python (MET CS 677) that I took during the summer was really interesting. I really hope we will continue our research collaboration in the future.
Can you share about your recent project on underwater shark detection? Were you able to apply concepts you are learning in your courses at BU MET?
This is one of the projects that I started to work on as a freelance IT consultant. It is linked to the Centre de Ressources et d’Appui (CRA), which is a research center at La Réunion (an island in the Indian Ocean) that studies how to deal with the shark attack risk there. This research center is mainly funded by the French state and operates in strong collaboration with the Université de la Réunion. There are a lot of shark attacks at La Réunion, and it killed the local economy. The goal of this research center is to unite different actors related to shark activity (biologist, politically responsible, researcher, or engineer, for example) to find non-destructive, non-invasive solutions that allow people to continue to swim in a designated area, while not hurting the sharks that are fully part of the ecosystem. Using underwater video analysis to detect sharks is one of their initiatives. The project will immerse cameras around a surf spot and have an algorithm that recognizes automatically when a shark passes in front. Images can be validated by an analyst on land, who can close the beach if needed until the shark leaves.
I entered the CRA as an intern in 2018 to work on a different project. My goal at that time was to develop an algorithm that detects, more generally, fish, and extracts the moment from the video, which was useful for biologists at the university who were trying to measure the evolution of La Réunion’s marine ecosystem—they had a huge number of videos to verify and not that much manpower. I developed and tested this algorithm during my internship and, at the end, the CRA offered to hire me as a consultant to study the possibility of developing a similar tool but for the detection of a shark. This is a way more challenging task because the images underwater are often unclear and it is sometimes very difficult to recognize a shark, even for a human eye.
So, I worked on that project for quite a long time on my own, learning thanks to online courses, and produced reports for the organization in the frame of my freelance mission. Arriving at BU MET, I was really interested in continuing that mission, so I talked with Professor Pinsky in order to organize a directed study on that subject. The project was very long because we started with a blank page. I had first to build and annotate a dataset, which we did thanks to the help of the Université de la Réunion and biologists at the CRA. Then we could start to build the algorithm.
My goal was to build a classifier using weekly labeled images (for each image the presence of a shark is known, but not its position); this is a tricky task because the images are very large and the shark can be very small or far away, but if we managed to do that it would reduce a lot the manual labor needed to do the labeling. After doing long research on existing works, I tried to implement a version of the Spatial Pyramid Network, which allows us to have a network resistant to scale changes, and we managed to have good results (around 80% success rate on the test set)—not enough for an application in the real world but still an interesting result to build from.
In order to compare our method with state-of-the art techniques for object detection, we labeled a part of the dataset with the position of the shark using a tool I built for the project, called a “lien label tool.” We could then try to train the network as Retina Net, which shows usually very good results of this task. The goal is to check if they can perform better and maybe to understand how we could use the strength of those techniques while using weekly labeled images (which is, in my opinion, a question that will take a lot of importance in the future of machine learning).
Many of the concepts I applied to that project were linked to my previous online courses in deep learning and computer vision, but I also used some concepts I learned at BU MET. The course IT Strategy and Management (MET CS 782) informed the launching part of the project and the organization of the labeling by operators, and the really interesting course Machine Learning (MET CS 767), with Dr. Eric Braude, pushed me to evaluate the use of new techniques for this project and helped me to structure my ideas. Of course, the fact I got to do the directed study and to meet often with Dr. Pinsky helped me greatly.
What are your next steps after graduation?
Because I want to be involved in research and continue to learn about deep learning and data analysis, I would like to start a PhD program where I can use what I learned during my master’s on a new subject. I started, in September, a short exchange program of five months at the University of British Columbia in Vancouver as a researcher, in order to work in genetics using deep learning and maybe start a PhD in bioinformatics after. I think that subject will tend to become significant over the next years and it is intellectually fascinating to work on that kind of data. I am sure that will be a wonderful experience.