Abstract: A triplet hierarchical metric network for 3D model sketch retrieval is proposed to address the problem that sketches are treated as ordinary images and their unique sparsity is ignored, and the intra-class differences between sketches and 3D models are not given enough attention, which affects the retrieval performance. Then, the network is fully constrained by multi-level joint loss across domains, so that the network learns to represent both single-domain intra-class differences and inter-domain relationships, which effectively improves the retrieval performance of the network. The experimental results show that the average retrieval accuracy of the proposed network on two publicly available datasets SHREC2013 and SHREC2014 is 87.7% and 83.3%, respectively, which is more than 0.5% and 1.5% better than the advanced work (the same base-net).