Research on Complexity of Single-Class Object Detection Model

摘要

At present, most of the object detection models are complex, mainly used to deal with large-scale, multi-classes object detection tasks. For medium-sized datasets, the single-class object detection task, it is unnecessary to use complex models. So we studied the complexity of the model. This paper tries two methods to reduce the complexity of the model: compress the model on the backbone network, and replace the original backbone network with a lightweight model, in single-class datasets, found on the detection performance in different complexity of the model presented similar convex function. It is proved that the medium complexity model is the best to solve the single-class object detection task according to the test results. Combined with feature visualization, and sensitivity analysis. It is proved that the medium-complexity model not only reduces the computational cost, but also has good generalization ability. For the datasets of different scales, the model needs to adopt different complexity. This paper provides a good performance model for single-class object detection tasks.