In the detection data set, the size distribution of sample instances is also very important. If the size distribution of different instances is concentrated in a small range, the detection model needs to learn very little size information, which can detect the instances more effectively; however, when the size distribution of the instances in the data set changes greatly, the model needs With multi-scale detection capabilities, this has higher requirements for the detector.

On the harness data set, the size of the instance is divided into two types:

● Semantic area. Semantic area is the actual area occupied by the instance in the image, expressed as the number of pixels in the polygon labeling frame;

●Bbox area. In the process of target detection, the definition of bbox is indispensable. When Mask RCNN proposes candidate regions, it uses the form of bbox. On the image, the bbox of the instance is the smallest rectangular bounding box covering the instance, and the area of the bbox is the area of the rectangular bounding box. Calculate the ratio of the different area of each instance to the image area on the training set and draw the corresponding histogram,

As shown in Figure 36. Among them, Figure 36(a) calculates the proportion of the semantic area of the sample instance to the image area, while Figure 36(b) calculates the proportion of the sample instance bbox area to the image area.