Research On Cross-domain 3D Object Retrieval

The rapid development of information technology has led to the wide application of 3D model technology,which is infiltrated into people's lives,such as 3D modeling,3D reconstruction,3D movies,3D somatosensory games and 3D medical.In particular,thanks to the rapid development and widespread application of 3D capture devices,a large number of 3D model datasets with diverse modalities have emerged.Every day,an exponential user-side 3D object data is uploaded to the Internet and the cloud.However,few of them own the complex and detailed label information.Most 3D object datasets,especially the use-side datasets have no labels which results in problems that is rush to be solved.For example,how to use labeled 3D objects to predict the category of an unlabeled 3D object,and how 3D objects from different datasets retrieve each other.Among them,cross-domain 3D model retrieval has become a significant and challenging task.In this paper,we address the case when 3D objects in target domain are unlabeled.For 3D object retrieval problems:(1)We propose a unified framework that joint deep feature learning and visual domain adaptation(VDA)which can be trained end-to-end.Specifically,benefiting from the superiority of deep learning networks.The framework can enable the statistical and geometric shift between domains to be minimized in an unsupervised manner,which is accomplished by preserving both the domain-shared and domain-specific features of each domain.Consequently,VDA can augment the discrimination of visual descriptors for cross-domain 3D object comparison,which is important for maintaining remarkable retrieval performance.(2)In order to search a 3D object in an end-to-end manner,we also propose an end-toend visual domain adaptation network for cross-domain 3D object retrieval(C3DOR-Net).It joints the domain alignment and feature learning in and end-toend manner,which avoid the error accumulation and improve the performance.The superiority of this method over the state-of-the-art methods has been demonstrated for two cross-domain protocols: 1)CAD-to-CAD object retrieval from two popular 3D datasets(NTU and PSB)in three designed cross-domain scenarios;and 2)SHREC'17 RGB-D-to-CAD object retrieval from the Object NN dataset that contains CAD and RGB-D objects;3)monocular image based 3D object retrieval.The comparison experiments show that the proposed method can significantly outperform the competing methods.