Figure 1.
A neural network architecture diagram is divided into three dashed sections labelled BackBone, Path Aggregation Network, and Output. The BackBone section contains three vertically connected BottleNeck Cross Stage Partial blocks followed by a Spatial Pyramid Pooling Fast block at the bottom. Arrows indicate feature flow downward through the BackBone. Outputs from the BackBone connect to the Path Aggregation Network section through multiple horizontal arrows. The Path Aggregation Network section contains a sequence of BottleNeck Cross Stage Partial blocks, convolution 1 multiplied by 1 layers, UpSample layers, convolution 3 multiplied by 3 stride 2 layers, and Concat layers arranged in alternating upward and downward feature paths. Arrows indicate bidirectional feature fusion through concatenation and upsampling operations. The upper branch contains a Concat layer connected to a BottleNeck Cross Stage Partial block followed by a convolution 3 multiplied by 3 stride 2 layer and another Concat layer. The middle branch contains a BottleNeck Cross Stage Partial block followed by a convolution 3 multiplied by 3 stride 2 layer and a Concat layer. The lower branch begins with a BottleNeck Cross Stage Partial block connected through convolution 1 multiplied by 1 and UpSample operations to higher branches. The Output section contains three convolution 1 multiplied by 1 layers aligned vertically, each connected to a corresponding BottleNeck Cross Stage Partial block from the Path Aggregation Network section, representing multi-scale outputs.Architecture of YOLOv5 (modified from Jocher, 2020)