Unsupervised Domain Adaption Of Object Detection Method

Object detection is an important technique in the field of computer vision,which automatically detects objects in images or videos and outlines them,and is widely used in autonomous driving,intelligent monitoring,and medical image analysis.With the further research on deep learning,object detection based on convolutional neural networks has developed rapidly in the past decade.Current object detection methods always assume that the training data of the model and the data obtained in the practical application are independently and identically distributed.However,under the influence of various uncontrollable factors in reality,this assumption is often not satisfied,which leads to a drastic decrease in the detection accuracy of the pre-trained model applied in practical.To solve the above problem,scholars have proposed unsupervised domain adaptation of object detection methods,which improve the detection performance of the model on the target domain dataset using labeled source domain datasets.This paper focus on the unsupervised domain adaptation of object detection problem in different situations of scene differences and proposes two different domain adaptation methods.The contributions of this paper are summarized as follows:(1)For the domain adaptation problem with small scene differences,an unsupervised domain adaptation of object detection method based on domain adversarial is proposed to achieve strong alignment of source domain and target domain features.Firstly,an fusion domain classifier with attention is proposed,which obtains multiple-scale feature maps in the network and combines them with attentional feature fusion to make it have stronger domain classification ability.Then,based on the fusion domain classifier,a set of domain adaptation strategies are implemented,including image-level domain adaptation,instance-level domain adaptation,and joint consensus domain adaptation,to improve the detection performance of the model and the robustness of the domain classifier.(2)For the domain adaptation problem with obvious scene differences,an unsupervised domain adaptation of object detection method based on style transfer and knowledge distillation is proposed.This method achieves domain alignment in a step-by-step way.Firstly,the Cycle GAN is used to generate intermediate domain data for input image-level domain adaptation.Then,the domain distribution information stored in the shallow batch normalization layer of the convolutional neural network are used to align the two domains in the network feature space.Finally,knowledge distillation technology is used for unsupervised learning of the target domain,enabling the network to learn instance-level domain knowledge.Experimental results show that the two methods achieve a m AP improvement of 0.3%～2.4%and 5.4%～5.7% respectively in domain adaptation experiments with similar scenes,and the second method achieves a m AP improvement of 0.2%～4.6% in domain adaptation experiments with obvious scene differences.The methods proposed can effectively improve the domain adaptation performance of the model and have certain theoretical and practical value.