Tutorial Series: Instance Segmentation on Vector's camera feed

Part 2: Labeling data with polygon bounding boxes

In our previous article, we explored the difference between object detection and instance segmentation. In this article, we examine how to efficiently construct a labelled dataset which will be used to build a model which allows the Vector robot identify another Vector robot using instance segmentation.

How to build a great labelled dataset?

Like all computer vision models, instance segmentation requires rich labelled (annotated) data for training the model. Specifically for instance segmentation, one needs to identify the contours of the object that needs to be detected in the picture. As you can imagine, drawing contours on a picture can be a task requiring a lot of human labor, and this is what made instance segmentation so hard and difficult to use.

Luckily, there now exist tools which can help make the job much easier for us. This article will explore how we can use one such SaaS tool called Roboflow. Roboflow supports a free limited user account and is a great service if you want …