Research On Unsupervised Domain Adaptation Based On Fine-Grained Feature Alignment

Modern deep learning systems have made remarkable success in many machine learning and computer vision problems.Their great success largely attributes to the presence of large-scale labeled datasets for learning representative features.An important assumption of machine learning algorithms represented by deep learning is that the data distributions are independent and identically distributed(i.i.d.).However,traditional deep learning models that trained on one dataset show poor generalization ability to another different but related dataset.This phenomenon is due to the presence of domain shift or dataset bias,while traditional machine learning algorithms require that the training and testing data should follow the i.i.d.assumption.To tackle the aforementioned problems,unsupervised domain adaptation serves as a very promising solution,which aims to transfer domain knowledge from one labeled source domain to a fully unlabeled target domain.Considering the empirical superiority of deep learning models,we focus on the deep domain adaptation problem in our paper.Firstly,we investigate the unsupervised domain adaptation problem in the context of image classification.The latest advances in unsupervised domain adaptation resort to discriminative knowledge transfer in virtue of pseudo-labeling to encourage the category-level feature alignment across different domains.However,these approaches are vulnerable to the error accumulation and thus incapable of preserving cross-domain semantic consistency,since the pseudo-labeling accuracy is not explicitly guaranteed.In this paper,we propose the Progressive Feature Alignment Network(PFAN)to align the discriminative representations between domains progressively and effectively,through exploiting the intra-class variation in the target domain.Specifically,we first devise an Easy-to-Hard Transfer Strategy(EHTS)and an Adaptive Prototype Alignment(APA)step to train our model iteratively and alternatively.Moreover,based on the observation that a good adaptation model usually requires a non-saturated source classifier,we introduce a simple yet efficient way to retard the convergence speed of the source classification loss by further involving a temperature variate into the softmax function.Experiments on three benchmark datasets reveal that the proposed PFAN exceeds the state-of-the-art methods.Secondly,we investigate the unsupervised domain adaptation problem in the context of object detection.The latest advances in cross-domain object detection have achieved remarkable results in virtue of adversarial feature adaptation to alleviate the domain disparity along the detection pipeline.Although adversarial adaptation is capable of strengthening the transferability of feature representations,the feature discriminability of object detectors remains less explored.Besides,transferability and discriminability may come at a contradiction in adversarial adaptation due to the complex combinations of objects and the distinct scene layouts across domains.In this paper,we propose a Hierarchical Transferability Calibration Network(HTCN)that hierarchically(local-region/image/instance)calibrates the transferability of feature representations for harmonizing transferability and discriminability.The proposed model consists of three components:(1)Importance Weighted Adversarial Training with input Interpolation(IWAT-I),which enhances the global discriminability via reweighting the interpolated image-level features;(2)Context-aware Instance-Level Alignment(CILA)module,which enhances the local discriminability by capturing the underlying complementary effect between the instance-level feature and the global context information for the instance-level feature alignment;(3)local feature masks that calibrate the local transferability to provide semantic guidance for the following discriminative pattern alignment.Experimental results reveal that HTCN significantly outperforms the state-of-the-art methods on several benchmark datasets.