Pre-defined references boxes of various shapes and sizes that a detector uses as a starting point to predict object locations.
Faster R-CNN innovation.
Helps models detect objects of different scales/ratios.