How to Use the Roboflow Dataset Health Check

Key Takeaways

hey there it's Joseph from Roma flow I'm gonna show you how you can use Robo flows data set health check to get the most out of your computer vision data sets now for today's example we're gonna be walking through the hardhat workers data set if you want to follow along you can actually find this data set on public dot Robo flow dot AI so if you go to public dot rubble fills out a I presented with all these free public image data sets and there's the hardhat workers data set I already have this data set in my account and I've pulled up the hardhat workers data set health check let's dive on in first things first I have seven thousand and thirty five images in my data set I don't have any missing annotations meaning all of my image files have a matching annotation file I also don't have any null examples 0 no example a null example is when I have an image that doesn't contain any of the objects that I wanted to detect in that given image so for example I don't have an image of a construction site that doesn't contain a worker a person or someone that doesn't have a hard head on I have twenty seven thousand and thirty nine annotations across my seven thousand images or approximately 3.8 per image that's good I mean it's pretty good feature richness across my three classes I also generally have pretty small images 0.17 megapixels my smallest one is point zero three megapixels my largest is 0.67 I'd actually click on that if I want to see this itty bitty small image and see its dimensions inside my dataset and here I see this this image here that's is 167 by 154 which is which is helpful okay now my class balance so having balanced classes is important in computer vision because we want our model to learn evenly across different objects that we want to teach it to recognize in this data set I might have some problems I have a pretty good representation of helmets nineteen thousand seven hundred forty-seven helmet examples I have about six thousand six hundred some examples of heads people without helmets but I only have six hundred fifteen examples of persons meaning people that don't have a helmet or a or just their face instead now I can zoom in and see show me Robo flow show me my data set every single image that contains a person remember there is 615 annotated people but that doesn't mean that there's 615 images of people why because there could be multiple people in a single image and in fact that's what we see here there's 209 images but a lot of these images have multiple people annotated like this one or actually this is a really good example here where I have all these people annotated and in the hard hat around those people as well okay now that's my health check so I also have here the size information me size matters because it helps inform the resize decision I want to make if I resize my images to square as most models require our my model is gonna be or our my image is gonna be stretched down or stretched up and it kind of depends in this data set it looks like my median average size is 500 by 333 so I might not want to go much bigger than 333 most the time you know we resize anywhere between 300 by 300 to 640 by 640 or somewhere in between sometimes bigger sometimes smaller it all depends of course on the context of your problem but I wouldn't want to go much bigger than 333 pixels here for the the height because that would stretch out my pixels more than I want the width is generally 500 pixels wide so maybe a good resize decision would be 300 by 300 perhaps a kind of depends again on the context to your problem I also see that the aspect ratio here a lot of my images are wider than they are square in fact you see here I have this line here where it goes directly across if an image is just as tall as it is wide that means it is a perfect square if it's wider than it is tall and it's a wide image and if it's taller than it is wide then it's a tall image I could have images that are very tall or very wide meaning if I stretch things to be square it might mess up their aspect ratio but conversely if I could preserve the aspect ratio it might create a lot of white or black padding in my resize decision rebel flow shows me those previews my pre-processing steps if I wanted to have a look so this is all useful information to inform my resize decision in particular now I also have the annotation heat map to understand generally in my image where are my objects appearing the helmets are generally across the my image the heads also generate across the top and persons are actually emanating from the bottom of my images that kind of makes sense right so this is like a gut check are my objects generally where I expect them to be across my images and if I were to make a resize decision or crop or change the size of my objects would they be in positions where they're in the frame of my image this helps make sure that you have a very visual quick check without manually combing through surfacing all these individual annotations in a quick one-stop way so that's kind of an overview of the things that are available to you and your health check now be sure to LIKE and subscribe to the Robo flow channel to learn more computer vision tips we cannot wait to see what you build using rubble flow to incorporate computer vision into your problems thanks so much for watching