 Our next presentation is Moses Leboer, Noel Jochem and Catherine Wilson, who are from the Conservation Engineering Group at the Alaska Fisheries Science and Centre of NOAA. They'll be talking to using machine learning to differentiate species and count fish in trawl nets. Hi, my name is Moses Lerber and today I'm going to be talking about how my team is developing automated video review tools using machine learning to help us differentiate between species and count fish in trawl nets. We're part of Alaska Fisheries Science and Centre's Conservation Engineering Group where we work closely with fishermen and other partners to address important industry issues like gear selectivity and energy efficiency. So why are we interested in automating video review? Well, first of all, manual video review takes up valuable time that could be spent fishing or pursuing other research objectives and also delays the time between data collection and data analysis. Right now, fishermen and scientists are collecting hours of footage from inside their nets, so being able to quickly review that data presents big benefits to both parties. And our team is working right now on developing a new net to reduce salmon bycatch in the $500 million Alaskan Pollock fishery. So being able to have a tool that helps us quickly review our videos would be very beneficial. Let's talk about our goals as we're pursuing an automated video review tool. First and foremost, we want a system that's easy to use and accessible to people with a wide range of familiarity with computer vision and machine learning technology. We also want a system that can handle the incredible variety of situations you can see inside of a trawl net. And of course, most importantly, we want a tool that will give us detailed, reliable, and standardized data when we're reviewing videos. Right now for our project, we have three main areas of focus, and those are data set development, object detection, and object tracking. Using Intel's open source annotation tool CVAT, we've been able to annotate over 5,000 images and 30,000 individual fish. And we've been using that data along with TensorFlow and Google Cloud to train and evaluate several object detection models. And right now, we're able to detect salmon and Pollock with a mean average precision of around 67%. And you can see some of our results on the right hand side of my slide. We're also in the process of implementing and evaluating deep learning based object trackers so that we can not only detect fish in our slides, but follow them and ultimately count them. Now counting fish in trawl nets presents a variety of unique challenges with video quality, the wide range of organisms present in the net, and high fish density. We've been able to address many of these challenges by training our models on large representative data sets. And we're particularly interested in the challenge of counting fish in red light footage. This is where we illuminate the net with red LEDs rather than white LEDs because this red light has been shown to minimize effects on fish behavior and get us better data. We've been training models with augmented data to mimic this red light scenario. And we've been getting encouraging results so far and it's something we're going to continue pursuing throughout this project. And the final challenge that we're all very familiar with by now is collaborating with teammates remotely. We found that GitHub and the ability to use virtual machines for model training on Google Cloud have been invaluable when collaborating with our teammates. Looking forward, we're striving towards a tool that has been developed with the input of both fishermen and scientists to make their jobs easier and facilitate innovation. Our automated video review tool will help us deliver data in real time, allowing fishermen to keep fishing and also saving scientists time. We also hope that our diverse data sets and established workflows that we developed during this project are going to allow us to answer future research questions more quickly than ever before. And with this, we're super excited to be working on this project, looking forward to learning more about your work and happy to answer any questions that you might have. Thank you. My question to you is, is I'll leave Matt to talk about the Pollock fishery and my I'm interested in a little bit and how much you've been held back or what's the story about getting the amount of funds you needed to run the types of tests you have on these open based systems which maybe the charges held you back from getting enough Processing time. Is that been a challenge? Where have you used most of your resources? Thank you. That's a great question. We have been lucky enough. Noah has developed a partnership with Google and their cloud computing resources. We've actually not had much of an issue getting the funding we need to run and train our models in Google cloud. Because Noah has been supporting that and as they've been developing their partnership with Google, we've gotten the opportunity to test out some of our model training without having to incur a large cost and also cloud computing and Google cloud and again not endorsing Google cloud in any way but cloud computing in general offers very affordable ways for us to train models quickly. We're lucky enough at FAO to have those types of relationships and these are the kind of questions in our collaborative suggestions that we're going to come up with at the end of how do we spread that love to other people working on these questions. Matt, can I pass the question to you? Thanks. Yeah, really interesting work. Where do you plan on taking this observation work? Do you see any room for say some kind of device on full nets? I think it's a marine mammal is detects or something like that. Do you know what I mean? I'm just trying to think around what further functions maybe do you perceive from your research work? Yeah, thank you Matt. That's a great question. We're working on developing a very flexible program that will allow us to detect any number of objects or organisms in a net. So right now we're really focused on salmon and pollock, but marine mammals is something that as long as we have a sufficient data set can certainly be implemented and easily combined to the project we're working on now. And long term, we're looking to develop a device that can help us analyze live feed and trauma. So that's certainly something that's on our mind for the future and we're really excited about. Anton's got a little question for you. I'm unmuted, it really helps. Are there plans to roll this out to this 500 million billion, what was it, industry? Or are you still in an experimental phase or when will that be field trials on real vessels? I wish I could give you a date as to when this program is going to be ready for all the fishermen out there in the Alaskan Pollock industry. Right now we're still in the development and testing phase. So we're continuing to fine tune our models and get this program up and running. If anyone is interested in getting updates about this project or learning about when we do actually deploy it, feel free to send me an email or email anyone on our team or shoot your email in the chat and I'll put you on our list. I'll finish with a question to you, Moses. What do you think your bigger challenge is? Is it technology or is it social or, you know, arranging for these types of tools to be rolled out? Is it the case you don't want to roll them out too early when they're not quite ready or, you know, where do you see your biggest hurdles to really be able to get this to scale? Well, you know, that's an interesting question. We're collaborating really closely with our partners and fishermen in the industry, so I don't see a big challenge in adoption in that respect because there is so much demand from fishermen for tools like this. I think the biggest challenge is getting a tool that is accurate enough that people trust it and trust it with millions and millions of dollars and their livelihood. I think that's one of the biggest challenges and something we're working really hard to address. Well, it's nice to have the problem of an industry with millions and millions of dollars and I hope that the breakthroughs that you do there and bringing the signal so much quicker back to the fishery is something that everyone can benefit from. So it's great to see your team work. Thank you very much, Moses.