Question: Anchors array about yolo-coreml-mpsnngraph HOT 4 CLOSED

These are the widths and heights of 5 "anchor" boxes (also known as prior boxes or default boxes). Recall that for every cell in the 13 x 13 grid we predict 5 bounding boxes? What we predict are actually widths and heights relative to these anchor boxes.

The reason YOLO uses these anchor boxes is as a hint to the neural network that most of the predictions will have one of these shapes. These particular anchor boxes were chosen by the authors of YOLO using clustering on the Pascal VOC dataset to find the most common object shapes. Using such anchor boxes is common in many object detection models.

from yolo-coreml-mpsnngraph.

It's constant for the training dataset used, in this case Pascal VOC. These anchors were calculated just once, before YOLO was trained. (The YOLO9000 paper explains how they did this in more detail.)

If you're training YOLO on a very different dataset with different shaped objects, you may need to calculate different anchors.

Turi Create, for example, also uses YOLO for object detection but has more and different anchors, presumably to deal with a larger variety of objects.

from yolo-coreml-mpsnngraph.

Is this constant value or somehow calculated? If the last then are you familiar with approaches to do so ?

from yolo-coreml-mpsnngraph.

Thanks, this is very useful information !

from yolo-coreml-mpsnngraph.