Research On Domain Adaptive 3D Model Retrieval Algorithm

Lacking labeled training data has long been a common challenge for applying deep learning methods to the 3D model retrieval task,and it leads to the insufficient generalization ability of existed algorithms.Domain adaption algorithm can learn shared knowledges and patterns from labeled domain and then apply them to unlabeled domains to solve such problem.By introducing domain adaption into the 3D model retrieval task,a domain adaptive 3D model retrieval algorithm is proposed in this thesis.Based on this algorithm,a high-performance 3D model retrieval system is built.Furtherly,the superiority of the proposed algorithm is verified by experiments.Firstly,inspired by the unsupervised domain adaption algorithm by backpropagation,a domain adaptive 3D model retrieval algorithm is proposed in this thesis.By introducing the gradient reversal layer and the domain classifier to the classic Stochastic Gradient Descent(SGD),this algorithm can learn the patterns from labeled data from one domain,and then apply them to unlabeled data from a similar domain.High classification accuracy is reached in the source domain,while very small amount of domain-relevant information is retained on feature vectors.Therefore,the proposed algorithm can achieve a similar classification accuracy in target domain as it does in source domain.Secondly,to satisfy the demand of large-scale data and real-time speed on the 3D model retrieval task,a high-performance 3D model retrieval system based on the distributed deployment technology and the proposed algorithm is built,and it can be applied to a variety of 3D model retrieval scenarios.Lastly,in order to evaluate the proposed algorithm,two popular 3D model datasets are chosen to carry out cross-domain 3D model retrieval experiments.The superiority of the proposed algorithm is verified by comparing its experiment results with cuttingedge domain-adaption methods and non-domain-adaption methods.At the same time,the function test is carried out on the proposed 3D model retrieval system.