Research On Unsupervised Domain Adaptive Object Detection

Object detection is one of the key research directions in the fields of computer vision,artificial intelligence,and so on.The main task of this direction is to identify and locate multiple interested objects in the image.The research results of object detection technology are widely used in the fields of face recognition,pedestrian detection,vehicle detection,and remote sensing detection,etc.Traditional object detection methods always assume that the training data and test data follow the same distribution.However,this assumption is hard to hold in the real world,because the source data and target data are usually collected in different environments.To this end,domain adaptive object detection is proposed to tackle the case where the training distribution differs from the testing distribution.However,another realistic problem is that,in many cases we cannot access the source domain data once the source model is trained due to data privacy and transmission issues,etc.This dissertation studies the problems of unsupervised domain adaptive object detection for these two situations.For the source-accessible problem,this dissertation carries out research work from two perspectives according to whether the source domain data is single-source data or multisource data.In the case of single-source data,this dissertation mainly studies how to ensure the semantic consistency in the adaptive process of feature space and how to deal with images and objects with different transferability to avoid “negative transfer”.In the case of multi-source data,this dissertation studies how to make full use of the knowledge of multiple source domains and how to fuse multiple knowledge to improve the detection accuracy of the target domain.For the source-free problem,this dissertation mainly studies how to use the implicit style information in the model and the framework of knowledge distillation to perform domain adaptation in the pre-trained source model's scenario instead of the source data for cross-domain adaptation.The main work and contributions are summarized as follows:(1)To address the problem of losing semantic information in the process of domain adaptation in single-source domain adaptive object detection,this dissertation proposes a cycle-consistent domain adaptive object detection network.Firstly,the features from the source domain are transformed to the target domain,and they are aligned with features from the target domain.At the same time,the target features are handled with similar operations.Then,a cycle-consistent loss is optimized to ensure that the semantic information is preserved before and after the style translations.Finally,the source domain features are equivalent to the source domain reconstruction features output by the source domain generator,and the same reconstruction operation is performed on the target domain feature to optimize the source domain and target domain generator.Experiments show that the method can effectively preserve the semantic information in the process of domain adaptation by aligning the transformed source domain and target domain feature and the feature of the source domain and transformed target domain.(2)To solve the “negative transfer” caused by directly aligning the images and objects of the source domain and the target domain in single-source domain adaptive object detection,this dissertation proposes a local-global domain adaptive object detection method based on an attention mechanism.Firstly,a global attention mechanism is proposed to highlight and weight transferable pictures to alleviate the negative transfer caused by direct improper global alignment.Then,strong matching between the two domains is realized in low-level feature spaces such as pixels and textures to alleviate cross-domain differences and preserve semantic information.Finally,for images with large domain differences,the attention mechanism is used to focus more attention on the object while ignoring the background information to improve the performance of the model.Experiments show that this method improves the domain adaptation process by treating different images and examples with different transferability and effectively alleviates domain adaptive object detection's “negative transfer” problem.(3)To make full use of multiple source domain knowledge to improve the detection performance of the target domain,a domain adaptive object detection method based on multiple source domain knowledge transfer is proposed.Firstly,the low-level features from multiple domains are aligned by learning a shallow feature extraction network to shorten the distance between the source and target domains.Then,each pair of the source domain and target domain is detected by using the subsequent multi-branch network.The high-level features of the standard domain are aligned to measure the transferability of each source domain to the target domain.And the target sample features output by multiple branches are fused based on this transferability during the test.Finally,image-level and instance-level attentions are used to promote positive cross-domain transfer and inhibit negative transfer.Experiments show that multiple source domain adaptation can not only improve the robustness of the model but also make full use of multiple source domain knowledge to improve the detection performance of the target domain model.(4)To solve the problem of source-data free domain adaptive object detection,in which only the pre-trained source model instead of the source data can be accessed.This dissertation proposes the source style transferred Mean Teacher for source-data free object detection.Firstly,the target domain features are transformed into source-like style features by using the batch normalization information from the pre-trained source model,which can make full use of the knowledge of the pre-trained source model.Then,the consistent regularization of the Mean Teacher network model is used to further distill knowledge from the source domain to the target domain.Finally,the robustness of domain-specific information is increased by adding perturbations associated with the domain distribution.Experiments show that by making full use of the pre-trained model's batch normalization information,the source domain's knowledge can be effectively transferred to the target domain.To sum up,through in-depth analysis of the technical bottlenecks faced by multiple problems of unsupervised domain adaptive object detection tasks,from the perspective of solving different issues,this dissertation respectively retains semantic information,automatic and dynamic weighted transferable pictures and objectives in the adaptive process,and proposes a multi-source domain adaptive knowledge transfer method.The research of source domain style transfer based on batch normalization is conducive to the development of target detection and has certain theoretical and application value.