 Going on to our next presentation now and it's by Angelo Mencarelli of Varkhanen University, a very active player in the development of this type of technology and I don't know if you'll talk to it but there's a very nice story happening in Holland with the recreational fishery. Anyway, he's talking to using computer vision to evaluate size and weight applied catch to enhance fully documented fishery. Hello, my name is Angelo Mencarelli and I work as a search for computer vision robotics at Varkhanen University and research in the Netherlands. Today I would like to present one of our projects. In this project one of the objectives is the development of a 3D camera system to classify count and estimated weight by species of by-catch as a result of artificial activity. Varkhanen University and research has the mission to explore the potential of nature to improve quality of life. In our campus we have a group on the ground for robotics that research develops computer vision and the agricultural robotic system for different sectors such as open field agriculture and productive horticulture fresh change and food livestock and of course marine. We are expert in artificial intelligence and sensing especially spectral machine learning and computer vision. Introduction of the lengthy obligation in Europe obliged fishing vessel to sort store and land all the undersized fish of the total allowable catch species that and that is a major addiction task on board in fishery catching and mixed composition species. Full document fishery aims to automatically registrate all the catches by species and their quantity making the distinction between catch over a catch under the legal minimum conservation side. Project provides the needed data and transparency without manually sort the undersized fish, store it and land it. It's what we call the transformation from a landing obligation to registration. Consulting consists of four partners, Varkhanen Marine Research, Varkhanen University, Varkhanen Plant Research and the Dutch Fishery Union Business. The project is founded by Lupino Maratmae and Fishery Found. In this project one of the purpose is the development of automatic detection system that is able to acquire images of the discard and classify the counts of the different species separating fish from the breeds and estimate the way of the discard by species. On the left side you can see the digital design of the 3D camera system that has to be mounted at the end of conveyor belt in the working area of the fish vessel. The system is equipped with a real sense camera that can capture color images and depth images using the stereo cameras. Actually we are using an industrial version of the Framos. On the right side the inner side of the 3D camera with the Framos and with the setup illumination that provides the fuselage. The basic concept is to acquire color and depth images of the discard together with the breeds to segment the fishes using the player thing and classify them, count and compute the volume of the fishes estimate weight. To test and acquire the images from the annotation training of the deep learning framework we mounted the camera on a mobile conveyor belt that simulated conveyor belt of vessel and we went several times in fish addictions and we acquired images of fresh discard and the breeds. After acquisition we manually separated the fish from the breeds and we weighed them. Using the color images a YOLO network identifies and classifies the fishes and counted them. A couple of examples of the preliminary results of the classification accounting process using the images of the onshore acquisition. Using the instant segmentation of the color images the depth images are segmented by a YOLO network and the volume of each fish is computed and weighed, the weight is estimated. Again an example, a preliminary result of the instant segmentation and the volume computation process. On the top left the color image and the right the depth image and the bottom left the results of the instant segmentation and the right the segmented depth image false color. From December 2020 we mounted three times the system on vessels for mechanical stress test, functional test and image acquisition with fishing activities. Two times for one week and once for 10 days. Here in the images you see the system mounted to the end of the conveyor belt in the working area of the vessel. Here a couple of examples of the preliminary results of the classification accounting process using images from the onboard acquisition. Although the preliminary results of the onshore and onboard test were positive more tests and training had to be performed in the next year and now the system has to be integrated into onboard fishery management. There are also other features that can be very interesting from the scientific point of view that can be computed. For example, the fish length and the system can be used to automatize the working process reducing the time that the light discard is out of the water. And with that I conclude my presentation. Thank you for your attention. Thank you very much Angelo. It's interesting to me that you're dealing with a very controlled environment there but also fish which are covered in debris mud very difficult often piled up so you have a number of stories to tell. I wonder if you've got any stories about the design of the overhead unit and how you've tried to overcome distance problems when you're trying to use stereo cameras, problems of keeping the plate clean, any of those stories or any cross fertilization because I know that Wachenen has a number of fingers and a number of pies when it comes to developing these technologies and I wondered if you could give us some insights into how your team's working and overcoming problems across projects. Thank you. Yes thank you for the question. Actually I'm lucky because I'm more a rotation so in principle when we can solve some problem let's say something mechanical we do with that. So that is the idea behind of this system. One of the basic problems that you have in deep learning in certain cases is the quality of image. So what we have usually in the open field of robotics for plant research where it's the group that come we try to let me say to protect the environment. So the first idea was to make a tunnel like we do usually for post harvesting or for that situation also in open field to have diffuse light to have no influence of the variation of the light. Second of course you know the problem of monitoring of the catching in the boat got also against some feeling that the fishermen has about the fact that they are in the view so they feel themselves let me say observed. So we were directly with the idea okay we want just to see the fish so we put something on the conveyor belt and so nobody is in the view already the fish and for us it's actually the best option because we are just working in the field of view of the camera with what we have and so the third point is actually the system we tested the system we say three times the last time for 10 days without cleaning the camera and the camera has little spots but we have also software to detect the spot so we know in the moment that in some points of the field of view of the camera this spot or the signal sense that that area is is blurred so it has to be cleaned and we have also in mind to mix this system automatically. Let's just stretching the problem a little bit here actually the main problem and I think that is the problem that everybody shall have in this situation is the illumination and not really the camera because you see we have a very cheap camera in principle so the illumination is the most difficult part and it's also the most expensive actually if you think that the whole system the most expensive part is the illumination purely for the sea it's really a big issue and the last and not least is the fact that we are using okay real sense camera but we are using a special way of course we use also the dick image for the volume calculation but actually we manipulate the software to not use let me say what you have usually a matrix image but a line-scan image so we use just 20 lines of the image we apply together and we have a very big image to let me say detect the large fish and of course also to avoid the problem of the pinhole effect of the camera so we don't want to have some deformation so if you have this image made in sort of line scan you have always the same let me say optical view of the image and you have let me say easier for some things the easier image for something else less but is in principle something that you can always have in the same condition and that's what you want I hope that I will be clear enough it is very clear and it's nice to get those insights because you know I think when people think about the problems that you're dealing with you start to hear threads between you and the dong about how illumination is at such a challenge and I don't think that would necessarily come to the first people's minds about what type of questions we could share amongst each other so that's great I'm going to pass you to Matt though and I'm sure Matt has some questions he's trying to design systems with equally you know capture size and so on and then start cameras as well and so Matt have you got any questions please Andrew? Great presentation Andrew I really enjoyed that and 3D and depth imagery is something that I've worked on for quite a long time as I mentioned yesterday did you did you think about using other frequencies that are available because something that I've found out is that because the IR spectrum isn't so visible underwater species haven't evolved to kind of have traits that some of the IR traits are more vivid in fact than the available wavelengths down at the blue end of the spectrum underwater so sometimes you can find in certain species they have really high characteristics to protect them and one option as well is filtering out visible light all together using an illuminator with specific wavelength with a filter on it so you have a totally pure well not totally pure it's because it never is but you thin down the illumination do you explore any of those options? Yeah we are exploring for other let me say projects actually this project is of course finalized to make this let me say this food of beneficiary for let me say a system that is possible to mount in the boat in such a way let me say of course if you are going in a perspective so you're going in infrared going near etc etc the information that you will have are enormous different situation and also situation where it's difficult to recognize some species because of course the color the RGB in the other frequencies it will be so clear to see and really also sometimes also to catch also the difference of something we don't know exactly I mean say now I'm saying something that is dangerous but you can see also fish that is healthy and not healthy for some situation and it will be let me say the next step really the point of course in the research site let me say you wanted to have okay again the illumination you want to have a very broad frequency for the illumination we are going to put a camera to analyze everything sometimes it's difficult because of course we are using a system that can be very hot on the on board it will be you need a dedicated let me say vessel that won't do this this kind of things but it will be very very interesting because the moment that you find that the narrow band that you need you you go to put the spectral camera it can be very cheap actually because you can just filter and it will be and with the right frequency and the right stuff yes definitely definitely in the future they we have to go in that direction also for other stuff we are talking now officially and this card but it's also for automatic selection and also for this card on the audition that you can do these kind of things and with the same concept actually and indeed it will be very very useful and of course another technology that's just kind of appeared over the past 10 years in robotics and cameras for robotics which is the the domain to look into for all of this of course isn't it is the time of flight cameras which use a specific wavelength single camera and produce a depth image at the same time I mean they're expensive but without having two cameras you get everything you need and they're specifically designed for that job aren't they yeah sometimes the problem is more than solution actually that you have done really the application but it indeed is it there's a lot of technology that in some fields can be really ported to the fishery of course the problem that we have in the fishery that is really corrosive environment and very difficult but let's say broadened thinking you have a lot of things that we can do really sometimes in very cheap way actually I'm just saying the last thing usually if you are concentrated only on deep learning and we are concentrating a lot about that you miss a very important part that part of the uncertainties of the deep learning are the point of the image if you miss this part you can have the best system of the world must still you have problems for last thing Kim I just want to touch on a point that Angelo brought up in his presentation which was consideration for the stakeholder and it kind of relates I wanted to ask earlier actually about this in the fantastic project that Norway is running out and I think it's something that we should think about in the forum is you know how what are the best ways to approach stakeholders because we can have all this wonderful technology but people don't like it it's not going to be adopted and I think having a think tank about you know how to incentivize use of this technology how to make it pleasant for people not invasive and be empathetic I think is a really important thing and almost as important as the technology having meaning for the stakeholders to actually have it with them and not and I think you highlighted something really important with a very complex piece of technology yeah we're going to get moving because thank you Matt and thank you Angelo for those insights that's very happy and we're going to be moving on to the Apple design for AI in the near future to make sure but obviously it's got to work you know people like things that work and then the next step is to make sure if it's working to make it palatable so this is where we're getting to and thank you Angelo for such a great presentation