Domain Adaptation For Image And Video Data

In the era of big data,to solve the problem of lacking labeled data,Domain Adaptation(DA)has attracted widespread attention.DA defines the difference between the data obtained in the changing scene and the original labeled data as the domain shift.By reducing/eliminating the domain shift,the problem of lacking labeled data could be solved by the robust classifier trained on the original labeled data.In this paper,the domain adaptation for image and video data in the real world is the focus of research.The real-world data are always high dimensional and with nonlinear structure,the dimensionality reduction of high dimensional data needs to be considered first.Then,the reduced dimensionality features are used for the unsupervised domain adaptation.In detail,the innovative research works are shown as follows:1.Dimensionality reduction methods based on local preserving projection assume that data have a linear structure,which is inconsistent with the data in practice.In order to deal with this issue,this paper proposes a novel nonlinear locality preserving projection method via a deep neural network,which replaces the linear projection with an appropriate deep neural network.In addition,motivated by the fact that the kernel technique can capture nonlinear similarity of features and help to improve separability between nearby data points,a locality preserving projection model based on Euler representation is proposed.2.Domain adaptation in Euclidean space is a challenging task on which researchers recently have made great progress.However,realistic data often does not exist in Euclidean.For example,many high-dimensional data in computer vision are in general modeled by a low-dimensional manifold.This prompts the demand of exploring domain adaptation between non-Euclidean manifold spaces.This paper is concerned with domain adaption over the classic Grassmann manifolds.An optimal transport based domain adaptation model on Grassmann manifolds has been proposed.The model implements the adaption between datasets by minimizing the Wasserstein distances between the projected source data and the target data on Grassmann manifolds.Four regularization terms are introduced to keep task-related consistency in the adaptation process.Furthermore,to reduce the computational cost,the model is simplified by preserving the necessary adaption property.3.With the massive growth of video data,the video domain adaptation problem becomes increasingly significant for practical tasks.Motivated by the excellent performance of Grassmann manifolds representation in video recognition tasks,we propose an optimal transport based video domain adaptation model on Grassmann manifolds.The proposed model reduces the discrepancy between different domains for the frame and video level features.First,the frame level discrepancy is reduced by extracting domain consistency features.At the video level,a fixed number of frame features are formed and represented as points on Grassmann manifolds.These points are fused with predicted labels to form fusion features.Finally,the video level discrepancy is reduced by minimizing the distribution discrepancy of the fusion features between two domains.4.Since Grassmann manifolds representation reaches excellent performance in video recognition tasks,it prompts to explore video domain adaptation between Grassmann manifolds space.However,existing Grassmann manifolds representation based methods represent videos as points on Grassmann manifolds,which ignores the implicit sequential information existing in the video data.In this paper,we propose a video domain adaptation model based on Grassmann manifolds trajectory representation to resolve this issue.The proposed model represents a video as a trajectory feature,which reserves the sequential information of the video.Moreover,to overcome the instability and complexity involved in Wasserstein Distance,the original Wasserstein Distance is replaced by the Max-Sliced Wasserstein Distance.