Tutorial Series: Instance Segmentation on Vector's camera feed

Part 3: Training a ML Model and doing inference

This is part 3 of our series on how to develop a Machine Learning (ML) model to allow one Vector robot identify the contour of another Vector robot in its camera feed (aka instance segmentation). In Part 1, we explored the difference between object detection and instance segmentation. In Part 2, we examined how to efficiently construct a labelled dataset which will be used to build a ML model that can do instance segmentation. In this part, we will take the big leap in training a ML model and then using it for inference to have Vector detect the contour of another Vector in its live camera feed. So buckle your seat belts and hop on for the ride.

Training a model on Roboflow

Once you have labelled data at Roboflow, it is relatively straightforward to train a new ML model to do instance segmentation. Roboflow provides you many tools to enhance the quality of your labelled dataset so that it is more ready and comprehensive. There are two broad steps: Preprocessing and Augmentation. Prepro…