Glenn Jocher: What is New in YOLO v5?

Key Takeaways

[Music] i think you started ultralytics for similar motivations to why we started roboflow which is the writing is is there for making these technologies accessible and usable for any variety of industries being able to train a custom detector that works successfully to find airplane parts to sushi detection to solving board games there's an infinite number of things that we could detect and understand it's a question of making sure that's accessible to everyday people and so now with that with that context i think that that's a phenomenal segue into this is why you've put out yolo v5 and so i do want to mention so that in your v4 you're credited with a notable acknowledgement because you introduced mosaic augmentation which again is an augmentation technique that takes advantage of almost like copying and pasting portions of image on top of each other it's almost like a cut mix but for object detection is maybe a fair way of describing it and so i would love to hear i mean so the motivations for yolova are accessibility ease of training high quality performance quick inference speeds i'd love to hear a bit about what you think yolo v5 does for accessibility that you think other models in the yolo family might not yet do and then a bit about what's different and what's new yeah sure so on the accessibility front let's see so i can give you a nice example of that so one of the issues that i saw so in my yolo v3 repo i just started working on it by myself and when more and more people started noticing it and they started raising their issues when they tried to train their custom datasets and i noticed that a lot of these custom data sets have objects which are not quite under the same i guess distribution of aspect ratios that you see in the coco dataset which is what the models are typically trained on so i realized that a lot of people were training models with anchors they're called that were designed for a different data set so i started working on code to automatically allow you to fit new anchors to your data i worked on this for a while and i introduced this using a k-means method which was the standard method at the time and then i followed it up with a second stage which was an evolutionary algorithm a genetic algorithm which would take the result of the key means anchors i use those as an initial starting point and then of all your anchors through up to several thousand generations on the actual cost function that the model is going to be trained on so rather than then use the k-means function which has i guess like a limited subset of cost functions this would allow you to actually evolve the anchors using the same types of losses that we're going to be seeing in training so i thought this would work better and it did work better to a certain extent but it actually wasn't enough because this manual step was a bit complicated people had to take the results paste them into their model configuration files sometimes the order would get a little screwy so there's a lot of places to break this method that i developed so in yolob5 i actually took this a step further something i call auto anchor and what this does is when you load up your custom data set you start training without having to do anything the code will look at your anchors and it'll compare them against your data and if they don't fit well there's a determination that they fall below a certain matching threshold then it will just start training new anchors automatically using the same method it'll get some k means and initial guesses and then it'll evolve new anchors using a genetic algorithm and then it'll turn around and automatically place these new anchors back into your model and train the model and save the model with these new anchors so there's not a single thing you need to do so if somebody shows up and they have one of these special data sets where they're fitting like long rectangles or vertical rectangles then hopefully the new repository will handle that a lot better much more automatically and much more seamlessly than anything in the past that i've seen really so this is just an example of like one of the small steps i've taken in many different aspects to try and increase the robustness and i guess the out of the box results you can say from someone coming in and using this for the first time that's fantastic that and i mean makes so much sense i remember yellow v3 struggling and actually i did hey means to determine where my anchor should be not even thinking about the or noticing the genetic technique that you'd published for making that even more robust [Music]