How to Capture and Label Training Data to Improve Object Detection Model Accuracy

[Music] hey everyone this is Evan from Edge electronics and this is my kitty cat Clio today we're going to show you how to capture and label images for training an object detection model right Cleo so by the end of this video you'll have a labeled image data set that you can use to train an object detection model on Frameworks like tensorflow as I show in my other videos the first step for creating any machine learning model is to build a data set for object detection models this means capturing and labeling hundreds to thousands of images of the objects you'd like to detect now it turns out that building a data set is actually the hardest and most important part of the process for training a model there's tons of details that should be considered as you're going around Gathering pictures but for this video I'll just show the basic process for taking pictures of your objects and labeling them using label image all right come on Cleo let's get started before you start Gathering images it's important to think about your object detector's application you want to think about the environment it will run in the perspective and distance it will see the objects from and even what type of camera it will use the pictures you take should look similar to what the object detection model will actually see when it's deployed in your application if it will be used in a variety of locations and conditions you'll need to get training images that represent those locations and conditions for this video the application I'm working on is a cat toy timer I just got Cleo a bunch of new toys and I want to set up a camera to watch Cleo as she plays with them it'll keep track of how much time she spends playing with each toy that way I'll know which one's her favorite so I can buy her more I'll work through the data Gathering process for this application as we go through the video alright so the first step is to start with the camera set up the camera so it's pointing at the area where you want to detect your objects now this step can vary in complexity depending on the scope of your application if you just want to detect objects in front of the camera you can make a simple setup where you set or hold objects in front of the camera at various angles and distances you can hold the camera and walk around taking pictures of the objects you want to detect or if your application is going to see the objects from many different perspectives and backgrounds then you'll need to set up the camera in a variety of ways to capture multiple views of your objects for my application I'll set my camera up to point at the floor where Clio usually plays with her toys now that your camera's set up let's take some pictures of the objects you can use any picture taking software that's installed on your computer if you've got opencv set up on your computer you can use a simple python script I wrote called picture taker it will open the camera and take a picture anytime you press p on the keyboard each picture will be saved into a specified folder instructions on how to install and use picture taker are available on my GitHub see the link in the video description below so the goal of taking pictures is to provide enough statistical data to the machine learning model for it to learn what your objects look like for an initial model you want to have at least 200 pictures in your data set you should have at least 50 examples of each object set objects in the camera's field of view and take pictures of them at different rotations distances and angles in my case I'm going to set the pet toys on the floor take a picture reposition the toys and repeat I'll change the camera's location height and angle a few times to get a different perspective of the scene if you know you'll be detecting objects and scenes with busy backgrounds you can also include other objects in the pictures or use various backgrounds to improve robustness and of course I need pictures of Clio too I'll let her play with the toys and take pictures while she's playing go get them Cleo [Music] once you've taken at least 200 pictures you should have enough to train an initial model if you really want to go crazy you can take 500 or even a thousand pictures production quality models actually would end up using ten thousand to a hundred thousand or even a million images so the more pictures that you take the more accurate your model is likely to be but make sure that there's enough visual difference between each picture you take or else these additional pictures that all look the same won't provide much value to the model okay we're finally done taking pictures now it's time for the fun part labeling them let's switch over to the computer where images are stored at this point you should have a folder full of pictures of your objects now we need to label the class and location of each object in every image the training algorithm will use this label data to teach the model what your objects look like and how to locate them in future images to annotate the images we'll use an open source labeling tool called label image label image has a simple and effective interface for drawing bounding boxes around objects and indicating their class label image is pretty easy to install if you're on Linux or Mac OS follow the instructions on the GitHub page for Windows label image is available as a standalone executable install it by going to the releases page linked in the video description below and downloading the windows file for the latest version next go to your downloads folder label image comes in a zip file that can be extracted to wherever you'd like on your computer I'll create a folder called label image on my desktop and then copy the folder inside the zip folder into the label image folder then go into the new folder and double-click the label image executable Windows might try and stop it from running if it does just click more info and then run anyway all right this is label image first open your folder of images by clicking the open directory icon and navigating to the folder where your images are stored click the select folder button and the first image in the folder will appear time to label some objects to start drawing a box click the create rect box button and click to draw a box around the first object in your image type in the name of the class and click ok to complete the label repeat the same process for each object in your image I'll give some more tips on good labeling practices later in this video foreign once you've labeled all the objects click the save icon and just save it as the default file name this creates an XML file inside the folder the XML file contains annotation data for the image the data is formatted in the Pascal VOC annotation format which is a standard way of storing bounding box coordinates classes and other information about objects in an image this format is commonly used with the tensorflow object detection API to train models if you're using a different framework the data can be converted to other formats using conversion tool like roboflow go check it out if you're interested okay back to labeling click the next image button to move on to the next image now let's learn some hotkeys to help speed up the process instead of clicking the create rectbox button you can just click press W on the keyboard to start a new Label Box instead of clicking on the save icon you can just press Ctrl s to save a file press D to move on to the next image or a to go to the previous image another way to speed things up is to pre-define the label map go to the folder where label image is installed and go into the data folder and open the predefined classes file delete the existing classes and type in your own classes in my case the classes are Kitty carrot Mouse fish and polka then save and exit the text file now go ahead and close and reopen label image go back to the directory where your images are stored now when you draw a Label Box the suggested classes will be the ones you put in the text file this makes it so you don't have to type out the whole class every time it also helps to prevent you from misspelling classes or using inconsistent class names which can cause errors or performance issues during training make sure to save before going on to the next image if you've made any changes all right let's go through a few more images and I'll give some labeling tips that will help your model achieve better performance object detection works best with objects that have clear and well-defined boundaries like a cat or a vehicle as such make sure to include the full object inside the bounding box it doesn't have to be perfect but err on the side of making the box too big to make sure that the whole object fixed Inside the Box it's okay if there's a few pixels between the box and the edge of the actual object you can also use the zoom tool to draw a more precise bounding box around the object so if the object is long and diagonally oriented in the image that means you'll get a lot of background when you draw a box around it but that's okay just make sure to have multiple examples of this object in multiple orientations in various images if an object is partially obscured like this carrot is just draw a box around the part that's visible a good way to think of how to draw bounding boxes for obscured objects is where would I want the model to predict this object's location wherever you want the model to predict the location that's where you should draw the bounding box it's also okay for bounding boxes to overlap so don't worry if the bounding boxes intersect each other however avoid using images that are too cluttered with objects here's an example picture of a crowd of people say you want to train your model to detect and count each person how would you label this image it's hard to Define where the edge of each person is as such your model will have a hard time predicting bounding boxes it's best to use images where there are clear boundaries around each object in some images it's hard to know what the bounding box for an object should look like or whether it should even be labeled for example I can barely see each end of the carrot past my cat's legs legs in this image should I label the carrot should I label the whole thing or should I label each end separately my general rule is this if I'm confused about how to label the image then I just throw it out the model learns best from Clear obvious examples if you have confusing or contradictory labels the training algorithm won't be able to minimize the model's loss here's another example of an image that I'll throw out because I'm confused if I should label the carrot or not the model will probably have a hard time telling that this small shaded object behind the cat is actually a carrot you want to start by giving your model simple examples that are easy to learn from after you get your model working well in Easy conditions then you can address more difficult visual situations by adding more images and difficult conditions to throw out an image just leave it unlabeled and then go to your folder and delete those images okay that's enough tips on labeling for now keep an eye on my channel and my website for more detailed tips on effective labeling strategies for now continue working through each image in your data set it's kind of a boring task so put on some music or enjoy a nice beverage while you work you'll know you're done when you click next image and nothing happens what a relief you can use this data set with my latest video showing how to train a custom tensorflow Lite object detection model with Google collab the video gives step-by-step instructions for training converting and deploying the model on a Raspberry Pi or other other device and the best part is you can do it all through your web browser and run training on a free collab server all right we did it at this point you should have a custom data set folder full of images and a label file for each image the next step is to use this custom data set to train an object detection model head over to my latest tensorflow video for step-by-step instructions on how to train a custom object detection model thanks for watching this video and we'll see you next time thank you [Music]

Channel: Edje Electronics

Views: 43,967

Keywords:

Id: v0ssiOY6cfg

Length: 13min 46sec (826 seconds)

Published: Mon Feb 27 2023