Transfer Learning for the Object Detection Model SSD

Abstract

The Single Shot MultiBox Detector (SSD) is a high performance object detection method.In general, a SSD model needs a huge amount of training data to build it.In this paper, we use a tranfer learning technique in order to expand the exsiting SSD model by using only a small data.In our setting, we have the 3-class SSD model trained using an enough data, and add new one class into the model. Our purpose is to do it by using a small training data. The SSD model is trained using three transfer learning methods with different ranges of freezing and initialization.Then, the accuracy of these models is compared. As a result, the accuracy of the model is not as good as a model created with enough data but is higher than a model trained with the same number of data. Therefore, although the proposed method can provide a certain degree of accuracy, it is concluded that another approach is necessary if the same accuracy is required as when using an enough data.