Research On Methods Of Cross Domain Object Detection

Object detection is one of the main research directions in the field of computer vision,and has wide applications in various fields such as transportation,finance,and national defense.However,in practical applications,the probability distribution of target samples often differs from that of the training set,which greatly affects the detection accuracy of the model on the target domain.One solution is to re-establish a dataset with a consistent distribution and conduct training,but this method is often costly and time-consuming.Therefore,studying how to improve the target detection performance in the absence of target domain labels is a highly practical issue.Existing cross-domain object detection methods still have some shortcomings,such as the lack of consideration for cross-domain performance differences of different samples during training,and the negative impact of domain-invariant feature extraction on detection performance.To address the above problems in cross-domain object detection,this thesis conducts research from three aspects,and the corresponding work and contributions are introduced as follows:(1)To address the feature alignment problem when the target domain labels are not available,this thesis proposes a cross-domain object detection method based on generalization error estimation on target domain.The method introduces a small number of auxiliary modules on the original detection network and estimates the model’s generalization ability in the target domain based on the inconsistency of the auxiliary network’s output.Based on this,the method directly optimizes the model’s generalization performance in the target domain.This method fully considers the differences between the classification subtask and the regression subtask in cross-domain training and further utilizes the pseudolabel strategy to fully utilize the source domain information.The experiments show that this method can effectively assist the detection model in unsupervised cross-domain training and improve the performance of the detection model in the target domain.(2)To address the problem of lacking consideration for the performance differences between different samples in cross-domain object detection,this thesis proposes a crossdomain object detection method based on hard sample mining.At the local level,this thesis uses the prediction results of the local domain classifier to discover the regions where the features are difficult to align,and calculates the corresponding residual attention value on the region to make the model focus more on the features in difficult regions.At the global level,this thesis uses the prediction results of the global domain classifier to discover the hard samples where the features are difficult to align,and calculates the corresponding residual attention to make the model focus on the foreground features in hard samples and ignore the background features.Experiments show that this method can effectively make the model focus more on the difficult samples and corresponding features in cross-domain training,thereby improving the detection performance in the target domain.(3)To address the ”negative transfer” problem in cross-domain object detection,this thesis proposes a meta-learning-based optimization method for cross-domain detection loss functions to alleviate the contradiction between the optimization of domain classifier and object detection head during training.Specifically,this method utilizes the metatraining and meta-testing phases to explore the mutual influence between loss functions of different tasks and force the gradient descent directions of different task loss functions to be consistent through the meta-optimization phase,which alleviates the ”negative optimization” problem caused by inappropriate feature alignment.Experiments show that this method can effectively coordinate the optimization contradictions between cross-domain tasks and detection tasks,and improve the detection performance of the model in the target domain.