Abstract: Sketches are advantageous in being easy to construct and unrestricted by language, discipline, age, and so forth, and the 3D model retrieval based on hand-drawn sketches has attracted increasing attention. However, due to the complexity of 3D models, intra-class diversity of 2D sketches, and the inter-domain differences between 3D models and 2D sketches, the sketch-based 3D model retrieval remains highly challenging currently. To address these issues, we proposed a 3D model retrieval for sketch based on adaptive multi-class centers and semi-heterogeneous network. First, the initial features of the sketches and the 3D models were extracted separately through two heterogeneous networks: a sketch feature embedding sub-network based on adaptive multi-class centers was designed to capture the intra-class diversity of sketches, and a 3D model feature embedding sub-network based on multi-view feature fusion was adopted to adapt to the complexity of 3D models. Then, using the label vectors with rich semantic information as guides, a homogeneous network was designed to realize the cross-domain shared feature embedding of the sketches and 3D models, so as to reduce the inter-domain differences. Comparative experiments on the large public data sets SHREC2013 and SHREC2014 demonstrate that the proposed algorithm is on par with or better than the state-of-the-art methods. 

Key words: sketch-based retrieval, 3D model retrieval, adaptive multi-class center, semi-heterogeneous network, semantic embedding