Creating videos of objects to be recognized

The ftc-ml tool uses videos instead of individual images because videos are an efficient package for managing potentially hundreds of images of the same object at slightly different angles/orientations and distances from the camera (poses). While video capture can be performed with any camera, it’s recommended that videos have exactly the same resolution as the camera being used on the robot. Ideally video capture should be done with the exact camera being used on the robot at the estimated height from the surface of the floor that the camera will be at. By using the exact camera on the robot, specific artifacts of the camera used – such as lens distortion and other optical effects – can be reflected in the training images which result in a much better overall object detection rate. Programs such as the Windows 10 Camera Application can be used to capture video from a webcam while it’s mounted on a robot and plugged into a laptop. It’s recommended to use the lowest frames per second (fps) setting possible, only because with a higher framerate the likelihood of getting multiple frames of the exact same image are incredibly high, and that’s just wasted frames that you have to label (or manually discard) and there’s no extra benefit in model training with duplicate frames (it takes longer to train your model). There are multiple web-based tools that will allow you to change the frame rate by removing frames from the video free online. For tips and best practices for creating the best poses go here.