Research On Domain Adaptive Object Detection

Object detection is one of the fundamental tasks in computer vision,which aims to locate and classify objects in images,and is widely used in fields such as intelligent monitoring and autonomous driving.Existing deep learning object detection methods are based on the assumption that the training and inference data satisfy the independent and identically distributed(i.i.d)hypothesis,but in practical applications,many scenarios do not meet this assumption,resulting in poor detector performance.Therefore,researchers have conducted studies on domain adaptation object detection,aiming to reduce the distribution difference between the source domain and the target domain through adversarial learning,which effectively improves the model’s generalization performance when enough data is available in the target domain.However,in practice,target domain data is often insufficient,resulting in poor performance of domain adaptive algorithms.Thus,this thesis studies a domain adaptive object detection method where the target domain data is partially missing.This method is essential in tackling domain adaptive detection tasks in practical scenarios.At the same time,domain adaptive object detection with partially missing target domain data is challenging,with the following obstacles: 1)Domain adaptive learning overfitting problem in the case of a sparse target domain.2)Feature distribution alignment issue under the inconsistent distribution information between the source domain and the target domain in the case of insufficient target domain information.To address the above difficulties,this thesis conducted the following research:(1)To solve the overfitting issue caused by the sparsity of target domain data,this thesis constructs a domain adaptive transfer module with input image self-augmentation and introduces semantic consistency loss.This module is able to effectively transform target domain images to the source domain,alleviate the overfitting phenomenon of the model on a small amount of target domain data,extract identifiable features in the target domain,and improve the detection ability of the model on partially missing images in the target domain.Compared with the baseline model,the proposed method improves the mAP by 5.7% and 2.0% on the Clipart1 k and Watercolor2 k datasets,respectively.(2)To address the feature alignment problem under partially missing target domain data,this thesis constructs a domain adaptive transfer module with deep feature semantic activation and designs a dual-branch domain alignment network.This module is able to enhance and transform the target domain features,reduce irrelevant background interference,and promote the alignment of local and global feature distributions.The proposed method achieved an 7.9% and 4.7% mAP improvement on the Clipart1 k and Watercolor2 k datasets,respectively,compared with the baseline model.(3)This thesis proposed a domain adaptive object detection method based on input and feature multi-layer consistency regularization,which jointly transforms the target domain to the source domain at the input and feature layers,and utilizes consistency regularization to maintain consistency in the transformation direction and promote accurate local and global feature domain alignment training.Compared to the latest DBGL method,the proposed method increases the mAP by 3.6% and 0.5% on the Clipart1 k and Watercolor2 k datasets,respectively.