Streamline your deal pipeline for animal science with airSlate SignNow

How to create outlook signature

to enjoy work between myself dr. John crawl Daniel Rubinstein Jason Homburg Tania burger wolf and Charles Stewart as a joint project between mr. Pond tech Institute Princeton University and the University of illinois-chicago so what do we mean by detection from an animal standpoint it's basically how can we identify this individual over time so from a mathematical standpoint we draw sift key points over the animal we put this image to a large search database and we can query on these key points to go and find the animal and see if we've seen in the past and this mathematical representation is very well understood and it actually ends up getting quite good matches if you build a large enough database and you haven't good enough search structure but to get to this point of having nice clear pictures of animals you have to pre-process given something like this so given this is your input you have to find a way to run a detection algorithm or a series of algorithms that can find all the animals that are unique figure out their species because you can't identify across species and you certainly can't identify by looking at different viewpoints so if you see an animal on the right side consistent over time you can't measure it against its left side they're not left-right symmetric so you have to have these basic you know annotation metadata to be able to solve the identification problem so what we propose today is a detection pipeline that actually gets you to that point so given an input image the first thing we will do is perform an image classification just to figure out what species exists in the image we then do annotation localization floating bounding boxes over all the animals we then try to figure out the species and viewpoint we perform a species specific background segmentation is trying to get rid of as much background vegetation or trees as possible and then we actually perform a novel concept called Aoi which stands for annotation of interest which tries to focus on what is the primary target of image right we've taken picture of a zebra we want to identify if we don't care about the 15 other background zebras that you can't really identify anyway to prioritize our focus so image classification again is trained to predict a multi-level medical inspector and it's generally considered as a high-pass filter to prevent irrelevant images from being processed things that don't contain animals I care about pictures of birthday parties things like that next thing as we go for annotation localization we actually find that in this domain if you have the right data set you can actually get fairly decent results in terms of high precision I recall on standalone images for of animals but when they start to herd and overlap and include each other you get problems our detectors based on a variant of Euro v1 by Redman and then from there we can take these bounding boxes and actually try to classify them this is this classic imagenet style you know give me a class for the individual that I saw be it a plain zebra or Grevy's zebra or it's a type of giraffe and actually go get the viewpoint we then perform that back in segmentation so it's trying to predict the course background segmentation mask and its task with basically providing a foreground wait for all the key points that we actually end up using by the sift matching algorithm this is trained with patches so we have no segmentation data at all which is a benefit and a curse we can actually train this patch wise and get fairly decent and even compelling results from a species by species specific standpoint from there we actually move on to Aoi classification as v pipeline component this is trained to predict an identifiability flag or an identifiable flag for each of the annotations that we see and this is tasked with eliminating as many incidental sightings of animals as that we scan this is generally meant to be a prioritize er generally you would want to go over and review all the annotations that you saw all the animals you saw to get the best identified population estimate but when you're starting to triage processing you have to start somewhere so we go for the most prominent animals in the image first this is trained with a input image and a mask for the annotations and you actually predict it yes or no as to whether or not it's prominent enough we also produce a new data set that actually includes six species of interests that we care about two species of zebra two species of giraffe well flukes and sea turtles and these are photographs taken by actual biologists wildlife rangers citizen scientists and conservationists in on the ground in Kenya or out the coast for the aquatic animals and we distribute this dataset freely you in the Pascal VLC dataset for anybody to actually try so these are harder and examples of Institute animals and it actually shows examples and scenarios that are passed what you'd normally see in an image net or Coco so in summary the idea is to reduce errors and the need for human in the loop reviewing for identifier notification we want to focus processing on the most identify all animals not any background animals that are just incidental and we wants to introduce a new data set for detection identification in this conformation this conservation domain for further results please come see them pupster thank you