The Design of Object Detection Network Based on Multidimensional Space Discrete Datum
Due to lack of design basis and could not be optimized for different datasets,the design of de-tection heads in target detection networks was addressed by proposing a corresponding design rationale through the study of a conflict rate model for the allocation of spatially discrete datum and positive sample points(groups)in dimensions ranging from 1 to 4.Based on the principle of location-based allocation,and consider-ing the distribution characteristics of positive sample points for a specific dataset in the spatial dimensions com-posed of four regression variables,a design pattern with the minimum number of points for detection heads was devised under the condition that the allocation conflict rate is not higher than the specified one.The objective is to reduce the imbalance between positive and negative samples and optimize computational resources.The process is essentially a regression benchmark encoding and positive/negative sample allocation design,aiming to achieve a balance between detection performance and resource consumption.The conclusion is that the re-gression benchmark based on the detection head of convolutional object detection network can use any combi-nation of four regressions,but it needs to be comprehensively designed in combination with the sample alloca-tion strategy and the maximum positive sample allocation conflict rate of the generated dataset.
target detectionsample allocation strategydiscrete position datumallocation conflict rate
万方数据
维普
