3.7 Non-max suppression
Non-maximum suppression, that is, only the maximum value (the prediction result with the highest probability) is retained.
There is also a problem with YOLO introduced earlier: there may be multiple prediction results for the same target. As shown below:
For the car on the right, there will be three grids that think the midpoint falls within themselves, so three predicted bounding boxes will appear when the result is displayed, which is obviously incorrect. This problem can be better solved by using non-maximum suppression.
The algorithm is as follows:
The result after non-maximum suppression is as follows:
Note: If there are multiple categories, you need to run non-maximum suppression once for each category.
3.8 Anchor Boxes
The YOLO introduced earlier cannot handle the problem of having multiple targets in a grid. To improve YOLO, the concept of anchor boxes is introduced.
Different from the previous grid corresponding to one result, each grid of YOLO introducing anchor boxes corresponds to multiple anchor boxes, each anchor box corresponds to a result, and the corresponding Y also changes, as shown in the following figure:
The midpoints of the woman and the car in the above image fall on the same grid, but are assigned to anchor box1 and anchor box2 according to their shape. Correspondingly, Y has also changed from 8 dimensions on the left to 16 dimensions (because there are two anchor boxes).
The specific algorithm description is as follows:
