Why are there only two detection heads(subnetworks) in YOLOv3 example?

In the original YOLOv3 architecture there are three detection heads(subnetworks), but in the object detection using YOLOv3 example of MATLAB why is there only two?

Answer 1

[Shubham]

Hi Akshatha,

The original YOLOv3 architecture has indeed three detection heads or subnetworks, which are responsible for detecting objects at three different scales. However, the example of object detection using YOLOv3 in MATLAB has only two detection heads.

The reason for this is that the MATLAB example uses a modified version of the YOLOv3 architecture called YOLOv3-tiny, which was specifically designed for constrained environments such as mobile devices. YOLOv3-tiny has only two detection heads, which detect objects at two different scales instead of three like the full YOLOv3 architecture. So it is not able to detect objects at small scales. However, it can still detect objects at medium and large scales.

Therefore, the MATLAB example has just two detection heads because it is based on the YOLOv3-tiny, not the full YOLOv3 architecture.