YOLOv3: An Incremental Improvement

Motivation

• Improvement of YOLO v1 and YOLO v2

Architecture

YOLOv3 makes detections at 3 different places in the network. These are layers 82nd, 94th and 106th layer.

• Network Input $(416,416)$
• For $82$nd layer the stride is $32$ and the output size is $13\times 13$ and it is responsible to detect large objects
• For $94$th layer the stride is $16$ and the output size is $26\times 26$ and it is responsible to detect medium objects
• For $106$nd layer the stride is $8$ and the output size is $52\times 52$ and it is responsible to detect small objects

To produce outputs YOLO v3 applies $1\times 1$ kernels at three output layers in the network. The shape of the kernels has its depth that is calculated by $b\times (5+c)$, where $b=3$ is the number of bonding boxes, and $c=80$ is the number of classes for COCO dataset. Therefore, the depth is $3\times (5+80)=255$, and the shapes of these output feature maps are $(13\times 13\times 255),(26\times 26\times 255),(52\times 52\times 255)$

Box Generation

Anchor Boxes is used to predict bounding boxes. YOLO v3 uses predefined bounding boxes to calculate real width and real height for predicted bounding boxes.

In total $9$ anchor boxes are used, $3$ anchor boxes for each scale. This means at each output layer every grid scale of feature map can predict $3$ bounding boxes using $3$ anchor boxes.

To calculate these anchor, YOLO v3 uses $k$-means clustering as in YOLO v2

• To predict center coordinates of bounding boxes $(b_x,b_y)$, YOLO v3 passes outputs $(t_x,t_y)$ through sigmoid function
• Then based on the above equations given in figure we get center coordinates and width and height of the bounding boxes.
• All the redundant bounding boxes are suppressed using Non-maximum Suppression