Research On Equivariance-Based Graph Data Augmentation Method And Its Application In 3D Human Pose Estimation

Graph data is a very common data type in daily life,and many other types of data can also be abstracted into graphs.Currently,researchers often face problems such as scarce real data and high annotation costs,and data augmentation is an effective solution.This type of method generates new data by mining the characteristics and rules of existing data,and uses the new data to train models to improve the performance of existing algorithms.Therefore,a very important issue in data augmentation is how to ensure that the generated data and the original data can satisfy the same underlying model.In other words,data augmentation should ensure that the predicted value corresponding to the generated sample is consistent with the predicted value corresponding to the original sample,otherwise,the generated sample is likely to become noise and degrade the predictive ability of the model.This requirement is more challenging since graph data is more abstract than images,text,etc.After investigation,the author found that the equivariance fits well with the requirement of data augmentation.Equivariance requires that the predicted value obtained after applying a certain transformation to the input data is the same as the result obtained by applying a certain transformation to the original predicted value.If the transformation on the input data is regarded as a generating transformation,then the features and label information of the augmented data can be obtained from the original data,and the data before and after augmentation can satisfy the same model.This thesis introduces equivariance into the process of graph data augmentation for the first time,not only theoretically deriving detailed formulas,but also proposing corresponding augmentation frameworks for graph node classification and 3D human pose estimation.The main contents of this thesis are as follows:1.Propose the graph data augmentation framework E-Aug.Graph data augmentation needs to ensure the consistency of the(feature-graph-label)triples of graph nodes before and after augmentation.This thesis refers to this property of graph data augmentation as augmentation fidelity.In order to ensure fidelity,this thesis modifies the requirement of equivariance for graph node classification problems and proposes “augmentation equivariance”.This thesis first assumes that the graph convolutional neural network(GCN)has augmentation equivariance for the transformations S on the input and T on the output,and then designs the generating transformation S and label transformation T based on the augmentation equivariance to obtain the augmented nodes and their corresponding labels.This thesis mathematically proves that GCN has augmentation equivariance for the two transformations proposed in this thesis,and proposes a new graph data augmentation framework E-Aug based on this.The augmented data obtained in this thesis can significantly improve the performance of GCN and graph attention network(GAT).2.Propose a 3D human pose augmentation framework MGPose.The 3D human pose is generally represented by the human skeleton,which includes the joints of the human body and the connection relationship between the joints,and naturally has a graph structure.This thesis combines the idea of autoencoders,regards the existing pose estimator as an autoencoder architecture,and proposes the MGPose framework.This thesis inputs the randomly masked features into the encoder to achieve data augmentation,and then reconstruct the joint features,and finally use the decoder to obtain the 3D human body pose.However,there is a gap between the masked data and the original data after passing through the same encoder,which may lead to inconsistent decoding of the human pose.In this thesis,the reconstruction transformation is carefully designed so that the decoder is equivariant to the transformation.The consistency of the masked data,original data and reconstructed data is guaranteed when they are decoded into 3D poses.At the same time,the parameter scale of MGPose is comparable to that of the original model,which is simple and efficient.