Events Add an event Speakers Talks Collections
 
MLconf Online 2020
November 6, 2020, Online
MLconf Online 2020
Request Q&A
MLconf Online 2020
From the conference
MLconf Online 2020
Request Q&A
Video
LINC: AI for the conservation of the African lion
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
144
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

Panthera leo (the African lion) is now an endangered species. Over the last 20 years approximately 42% of its habitat has been lost, creating fragmented populations across the African continent. The accurate monitoring of lion populations and a better understanding of the connectivity between them is critical to maintaining the genetic viability of these increasingly isolated populations.

The Lion Identification Network of Collaborators (LINC) is an open-source platform designed to change the base methodologies of how lion research can be enacted within these fragmented geographies and diverse conservation efforts. LINC employs a custom web application, innovative AI tools and a collaborative database allowing the consolidation and retrieval of lion data by conservationists, researchers and government wildlife management. LINC also provides a platform for social interaction and data sharing between conservation efforts and government institutions, shaping and informing conservation policy. The LINC project has built this foundation with ten partner organizations, KWS (Kenyan Wildlife Service) and support from Microsoft and the National Geographic Society. This strong interlinked research community enables conservationists and decision-makers to pinpoint priority areas nationally and internationally.

The workflow of the LINC system is as follows: conservationist capture in field images of individual lions which they transfer into the community database. Once the images are entered in the system, identification is performed using a set of ML techniques. The first technique uses facial features, while the other employs a whisker spot matching method. The predictions are returned to the conservationists via the UI who then correlates the appropriate metadata and informs on the ground teams of the lion movements.

This talk will focus on the development of the Machine Learning (ML) models for identifying individual lions across varied image sets, reducing the time and human resource needed to utilize large data collections. While human face identification has been an active (and sometimes controversial) field of research, the case of Computer Vision (CV) for unique individual identification for different animal species has largely not been tackled by the research community.

First, we will go over the challenges of the unique data set of african lions, and cover how deep learning techniques are used to identify unique individuals across time using facial features. Later, we will talk about the process of whisker spot pattern matching - a technique widely applied in field by conservationists since the 1960's, and still the dominant method today. We will dive into several techniques for automating the pattern matching, and speak about some promising results. Lastly, we will suggest research directions for the future, that could also be of use for identification of other species.

About speaker

Agustín Mautone
Machine Learning Engineer at Tryolabs

Agustin is a machine learning engineer at Tryolabs with a special focus on computer vision. Over the past years, Agustin has been part of several projects, ranging from lion face detection and identification, to image inpainting using GANs. Currently, he is working on a challenging CV project that highly accelerates the production line for one of the biggest retailer companies in the US. Also, he oftentimes participates in the local community sharing knowledge and experiences about machine learning projects in general.

View the profile
Share

Hello, everyone and welcome back. We now have a ghost in my town joining us. He is actually calling in from Uruguay Augustine, please join us. Hey, hello, and welcome everyone. Today. I'll be thinking about you think and how you say I for the conservation of the African Lion. So my name of the older than ancient is everything else on it. I work in Junior and that work. So a little bit about Trial of soda machine, learning consultant, a company, which partners with companies to bring business value and help them throughout all the

stuff from the very Inception. You are indeed to put in those systems into production. Let me tell you a bit about the agenda today, and what's their mission and then I'll go first of the phases of the Lions and then we're going to a two-faced identification. So we're using basis of information and then we source of information. So, okay. Let me introduce you to me. So, where does link exists African lions are endangered species of? You might guess so over the last 10 years

and under activities as a consequence on how the populations get increasingly fragmented and lions have the wrong over bacterias, that no one research group activities throughout their entire range of their movements are truly understood throughout Africa. So, how do you think Nine Star Trek at the moment? So it's hard to identify. So do use this GPS colors that you can see here, which are like everyone to three years. So it's super efficient. And the most powerful part is Appliance have to be sedated to to put

color into their neck. So that's very harmful for 10. So we're trying to avoid that. This is what links in fixing link Stanford. Is this a Custom Fabrication? Researchers and government wildlife management, but how do you think they identified him using images? Cuz I just said that it was not an easy task, at least. So do you stay with your spot patterns? You can see that they use this chart which contains to roast first. There is the reference wrote which corresponds to the most complete row of whiskers at

the top. And then we have the identification road, which contains a complete row with only four spots. Then they fill out this chart, which looks like they're related position of the identification spots in relation to get side. Unique identifier for every lion super slow. And they don't want to do this process over 400. So it's data set of images of lines in the world with bounding boxes around, their body, and body parts so that they can't we can do the object detection Park and then we

also got identification bit. Which one is this information to build a motor that they will talk to make this process and make it faster for consideration is to identify the Lions. So this is where we partner with them and gave it a dis plan of action. So first, we will get the data and try to understand any potential Folsom negative bias is very serious for us to go to swear when I mean whether the project is viable or not. And whether we need to keep working on the date,

we will go into a free Facebook from the first case would involve using object detection of the faces and body parts of the lion just to propose is that you don't want to face an identification card. So the first one would be, I didn't get to get the lines using the, their faces because we decided to begin with this part because images were allowed to wear higher quality. Is there a year? So there's a lot of information Network identification methods because it's what they use

to identify the line and we thought that it might contain some potential. Just getting to the brute look like to see the images that contain the Worcester area with a single bound books for each week. So there's one guy found in box around each of their whiskers. And then on the left, we have the alliance in the world. This is quite interesting because it's got levels around each of their body parts, but also it contains orientation for me to break simple. If you look at

this little Theory, which means he'll and then pass to the blinds look into the front end. For example, in this case. If you look it says t v. E r e r. Because it's not very common enough, attention problems, but still if you look at the car, but it does not contain a label of sale. So we don't want to go. So we had to go through to get to a point where was consistent and coherent world, all the decent. And after that we came up to us to move on to the modeling face. So

we decided to go with Foster r-cnn, which is the network. We have the most experience and ability to work to work with his day. Job. We got some awesome results. You show results for Luke and Brady. Sure you can send example of a little sleep fine. But so we have to take a careful, look at our code and find where they issue was happening for. The first one was the labels contain orientation information in when we Daytona Edition for this date. I said we were using

this information of political changes. So unusual object detection, for example, if you're trying to say Play labels. The right the right becomes the left and the last becomes the right if otherwise you're given conflicting information. So we had to build our custom to take into consideration. All these jerks which prevents the labels from the same class to be over in like so in this case we have to label for Tempo, but they're not giving removed from his head, right? And the other hand from. So

we have to build our own custom animal shelter to take into consideration. We finally got some really good results. So this was great, but still we wanted to know whether the most natural repeated to the to the datacard what happened to us. Since we only have four vials of the day. We wanted to make sure that they had in their only after they don't like we wanted to make sure that it was the old as well. So we came up with this this year, was the year that the Lion King came up at the cinemas. So we decided to try it on some of the

scenes of the movie in the movie. It was working great. And there was no way that the murderer was over. Feed it to this last piece. They don't even text you so much already. Obviously the most machine learning is very hard to get all the cases. Grab us an Omission. For example, when they get too close from each other,. Not since it's hard to generalize Welch, all these cases, but thankfully the same techniques with in for the body part section. We got some results. The only issue here. Is that a good resolution that? They are very

pixelated. Sometimes, it's hard to get to know whether or not you weren't. So we decided it was good enough to move on to the next place. So we going to see identification of faces. Okay. So awesome mansion identification face working with the face later because we thought it was a bit more feature-rich and higher-quality. But still it was going to be hard to get it. Always the correct line. So I took one, like always points exactly, which lion is confident that if we tend to tiny lions to the conservationists, we were going to be in

till at least 5 to tell which they can go later. We should still amongst the breakdown savings since they don't have to go over the $400 that they have on their database and the other, every time we would have to have a new class to the classifier, which would need to be trained to predict that life. So we don't want that. So we decided to use another approach which I was going over. So did they looked a bit like this? So we have crops of the phases of the Lion in their habitats. So the

container / 359 and over 6,000 images, images for lion, Bargain, Outlet Holdings account. So, we didn't want that to harm our training. So okay, for the image processing, we got inspired by a lot of dogs. How to identify a whale images to try to identify Wales based on the images of their tails when they come up out of the water. So one of the parts that goes into the most inspired this, for example, first base on something, that's very common on Chemung Place identification, which is trying to align their

eyes of the individual. Exactly. On the same location for every month. So in this case, we would use the architecture to try to protect the eyes of the of the Lion in August 2nd to be exact. Your time identifying which features of the face are the best to try to identify an interview. So this was very helpful. And also what time does conservationist we found up the volume, two spaces changes a lot over there. So for example in this Alliance don't have them when they grow older

and their heads. So what we decided to do, what you see this property damage of the right part of the bass lines throughout the day, they were ages. So so Did you go to to the Moline case? Okay, for the moment, we decided to go with the restaurant 50, which we trained as a pacifier at the beginning. So we would get a single class for each client. And once we would remove the top that yours. So we wouldn't expose at 512 and bidding which we were going to use to try to generate abstractions images that we are receiving. A lion face

has been compared to try to guess whether it's which lion is the one game picture. So this works and if we get three in bed or Temple and two in bangs are super close. On one is super far. We know that there are things that are super close have higher probability of being from the one that's super far. From using this technique. We can then go and retrieve the closest neighbor K nearest neighbor. So that's what with the objective I was talking about what I did again and

we will go over there and throw those. And then since you're there and then we would just get the face of the lion. Listen to Jerry Peterson been, in this way of working suspicious suspicious late. We got 95% accuracy, very fast, like that, much work. So we had to take a better. Look at the data to find out. Why. so high without Taking that much like training for a gas station or something. So you can see images of two lions and part-time police case you going to see that this is not

the same from this one, but it's very close. And for example, DC much is not the same from this one, but it's very close. So you might be wondering what say, you should their solution is that when we split the data into trainin test, we will get some of those images into the tested and some of those images into the Chinese it. So when we were trying to get evaluated with on the desk is almost exactly the same as the so we were like pretty. So so we decided to which we have Trainer

and also. Do Slimes something very different images from each other. And that way we could ever do we go to school. Going to see an example of the query images. This column contested that we cancel the database and their Bros. Digital is returned or orange and 85% of which we were very fond of because we return you said we were to get get over 95 or higher accuracy. So this was really good for us. We were very happy with it, but then we decided to move on to

the We were only, we only have two weeks left. And this was for the conservationist was too helpful already. So I think we only had two weeks to work on this project. We decided to do a quick way to get to understand, whether the problem was worth investing and spending some more time after the road. So when we go to start with the problem, we face it, two different approaches. So the first one would be using a pattern matching and then just trying to match point Cloud

which we would extract from the images by using the object detector, to try to get the centers of each of us in this pointclub which week and then try to match to other So, if we throw on some research, has a lot of points in the use of tournament, so we decided to go with no nuts. Cashier, in Point Break, which is, which applies to generate a match between the two sets of coins until you can see an example where it's working. Right? But this was not always the case here. We going to see a place

where he fails because the arrow from the detector go. You can see the blue sets of darts container from the three rows of winter. While there, I thought contains four sets of three. So we we had, we still think that this method contains a lot of potential, but we did not have enough time to go into it a lot because the, our time was running out, but we got in contact with the academy and we started sponsoring to students which are currently working on

this problem with using this point of metals and images. Are you saying they want to try to match the matching pieces of this student is ongoing? It is going to be finished by which we are. And then we think that we can move on to the mixture of both. Which is the best way to approach this problem, and I think we only have a few minutes left. Maybe if you have any questions, go ahead and thank you. August and amazing presentation. We are getting tons of Kudos. Everyone sticking this super-cool. I don't see any questions coming through, but I see a lot of Applause for you and credible research. What?

What? Tasks you had done very inspiring, right? It makes the day go so fast and we have our next speaker standing by. So let's thank you so much Augustine, and we love this presentation. And again, you can put it in, you know, the backstage chat on the stage chat here and he'll keep monitoring through the day.

Cackle comments for the website

Buy this talk

Access to the talk “LINC: AI for the conservation of the African lion”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “MLconf Online 2020”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Artificial Intelligence and Machine Learning”?

You might be interested in videos from this event

February 4 - 5, 2021
Online
26
104
ai, application, bot, chatbot, conversation, data, design, healthcare, ml

Similar talks

Somnath Banerjee
Head of Shopping Discovery at Pinterest
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Vinay Prabhu
Chief Scientist at UnifyID
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Meghana Ravikumar
Machine Learning Engineer at SigOpt
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “LINC: AI for the conservation of the African lion”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
949 conferences
37757 speakers
14408 hours of content