Machine Learning Toolchain

Creating custom TensorFlow Detection Models

Resources

• FTC Machine Learning Docs

Setting up the Account

To utilize the FTC Machine Learning feature, you need to have an account and be paired with a team on the FTC site. Then you can go here, to login to the ML toolchain.

Creating a Video of the Object

Your first step, is to create a video of the desired object. Due to the fact that there are many frames to analyze in a video, the FTC-ML Toolchain uses videos instead of images to recognize the image. Ideally, you should be using the camera used on your robot to take the videos from about the height that it will be on your robot, this ensures highest accuracy.

Video Upload

Now you need to upload your video to the ML software. On the FTC-ML dashboard, hit the upload video file, and select the video file from your computer. Wait some time, and the video will upload and the frames will be extracted and appear in the Contents Tab. After fully loading, the video description will become a link, click on the link to get the next step.

Labeling Frames

Now, you need to physically label the object in each frame of the video. However, the ML tools will automatically label sequential frames for you, however you must supervise. For information on all the controls and how to operate the labeling, look here.

Creating Dataset

Click the checkbox that is next to the video and choose the amount of frames for training and testing. Usually in ML models in general, 80% are used for training and 20% are used for testing the accuracy. Then you need to configure some things for the training, all information that is needed after the creation of the dataset can be found here.