S<SUP>2</SUP>Mix: Style and Semantic Mix for cross-domain 3D model retrieval

作     者：Fu, Xinwei Song, Dan Yang, Yue Zhang, Yuyi Wang, Bo 

作者机构：Tianjin Univ Sch Elect & Informat Engn Tianjin Peoples R China Hisense State Key Lab Digital Multimedia Technol Qingdao Peoples R China 

出 版 物：《JOURNAL OF VISUAL COMMUNICATION AND IMAGE REPRESENTATION》 (J Visual Commun Image Represent)

年 卷 期：2025年第107卷

核心收录：

学科分类：0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 0835[工学-软件工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Tianjin Municipal Natural Science Foundation, China [23JCQNJC01980] National Nature Science Foundation of China China Postdoctoral Science Foundation, China [2021T140511, 2020M680884] 

主　　题：3D model retrieval Image style transfer Unsupervised domain adaptation 

摘      要：With the development of deep neural networks and image processing technology, cross-domain 3D model retrieval algorithms based on 2D images have attracted much attention, utilizing visual information from labeled 2D images to assist in processing unlabeled 3D models. Existing unsupervised cross-domain 3D model retrieval algorithm use domain adaptation to narrow the modality gap between 2D images and 3D models. However, these methods overlook specific style visual information between different domains of 2D images and 3D models, which is crucial for reducing the domain distribution discrepancy. To address this issue, this paper proposes a Style and Semantic Mix (S2Mix) network for cross-domain 3D model retrieval, which fuses style visual information and semantic consistency features between different domains. Specifically, we design a style mix module to perform on shallow feature maps that are closer to the input data, learning 2D image and 3D model features with intermediate domain mixed style to narrow the domain distribution discrepancy. In addition, in order to improve the semantic prediction accuracy of unlabeled samples, a semantic mix module is also designed to operate on deep features, fusing features from reliable unlabeled 3D model and 2D image samples with semantic consistency. Our experiments demonstrate the effectiveness of the proposed S2Mixon two commonly-used cross-domain 3D model retrieval datasets MI3DOR-1 and MI3DOR-2.