Bounding Boxes

How it Works

Bounding box localizes the boundaries of the object, i.a. class, defining  highest and lowest points of the object on the image with a simplest annotation in the shape of rectangle shape. Imagine tracking people in the supermarket security camera?

 Challenges and solutions:

 *   Models  do not perform well when the object is too wide.  For most familiar use cases – street view camera – cars and pedestrians are boxes easily. That’s why there are so many pre-trained models available for these industries YOLO.  But take industrial facility, imagine to annotate into the box a scrap metal materials on assembly line. Instead of “localization” of the metal particle, the box instead will capture the whole line, because of the object’s shape.  In such cases, in TaQadam we use COCO format to allow machine learning engineers to use the appropriate shape in class training.  We also have features to allow to identify object as one box even though it is split on the image into segments.

 *   Instances, multi-tagging systems allowing to associate the same box with several layers of attributes. Our Platform in TaQadam allows flexibility in designing the annotation process to allow build more than one model.