Deep Feature Learning For Data Irregular Distribution

Image is an important source of information for humans.With the eapid advancement of technologies,such as the Internet,smart phones,and social media,people can easily obtain a large number of image resources.Different from the single small-scale academic dataset,the large-scale dataset often shows multi-model distribution,long-tailed distribution or even unlabeled data.We call this kind of data as data irregular distribution.Irregular data distribution widely exists in academic research and real world.Moreover,with the rapid development of deep learning in recent years,it manifests strong capacity in solving the image feature learning and has attracted great attention in the research community,Deep learning has been widely used in the academia and industry.Combined with the above two points,Deep feature learning for irregular data distribution has significant theoretical research value and strong practical application needs.This paper studies three common problems of data irregular distribution,that is,the long-tailed data distribution,the multi-modality data distribution,and the misalignment between source domain and target domain.1)To address the long-tailed visual recognition,This paper proposes two methods,i.e.,a learnable data expansion for tail data,called Feature Cloud and improving long-tailed visual recognition with model jitter,called MBJ.This paper proposes to transfer the rich intra-class diversity of the head class to the tail class for alleviating the insufficient intra-class diversity of the tail class.To this end,this paper proposes to expand the distribution of tail class in embedding space,which aims to make the tail class having the similar distribution to that of head class.Specifically,a tail feature is replaced by a cluster of feature which is called Feature Cloud.The distribution of Feature Cloud is learned from the head class.The experimental results show that Feature Cloud has achieved the state of the art performance on long-tailed representation learning and classification learning tasks.This paper proposes to enrich the tail data by the jitter information between historical models.With the training iteration,the parameters will continue to change,resulting in the weight jitter.Correspondingly,given an image,the feature vectors generated by the two historical models are also different,that is,the feature jitter.This paper points that the tail data can benefit from these jitter information.To this end,we use memory bank to collect these jitter information for providing the additional diversity for tail data.The experimental results show that MBJ has achieved the state of the art performance on long-tailed representation learning and classification learning tasks.2)To address the mulit-modality data distribution,this paper proposes a Memory-Augmented Unidirectional Metrics,called MAUM.MAUM consists of two novel components,i.e.,unidirectional metrics and memory-based augmentation.Specifically,MAUM first learns modality-specific proxies independently under each modality.Afterwards,MAUM uses the already-learned MS-Proxies as the static references for pulling close the features in the counterpart modality.The crossmodality association is further enhanced by storing the MS-Proxies into memory banks to increase the reference diversity.Importantly,we show that MAUM not only improves cross-modality retrieval under the modality-balanced setting,but also gains extra robustness against the modality-imbalance problem.3)To address the misalignment between the source domain and target domain,this paper proposes an unsupervised domain adaptation method for multi-domain image style transfer,called IPGAN.First,IPGAN translate the styles of images from the source domain to the target camera domains and generate many images with styles of target camera domains.Moreover IPGAN ensures that the transformed image has the identity information consistent with the original image.Next,the source domain data with annotation information after style transfer is used to train the model and applied to the target domain for testing.The experimental results show that IPGAN has achieved very competitive performance.