Rethinking sketch-based 3D shape retrieval: A simple baseline and benchmark reconstruction

Shuang Liang, Weidong Dai, Changmao Cheng, Yichen Wei, Yiyang Cai

Abstract: Sketch-based 3D shape retrieval has been receiving more and more attention. However, most previous works have two important problems: complex pipelines of their methods and unreasonable data split of the public benchmarks. The growing complexity of previous works leads to the difficulties in analyzing and following them. The unreasonable benchmarks cause the information leak of the evaluation of algorithms. In this paper, our work aims to addresses the aforementioned two problems for sketch-based 3D shape retrieval. First, contrary to the previous complex approaches, we provide a simple and effective baseline method based on prototype learning. It outperforms all the previous methods by a clear margin, indicating that the complexity in existing literature might be excessive. Second, leveraging our method for analysis, we demonstrate that the current benchmark is problematic in its data split, which impairs its validity for evaluation. Accordingly, we reconstruct the benchmark via more rigorous data splitting schemes based on zero-shot learning task. It shows that our prototype learning-based method shows more significant advantages over previous methods when facing unseen classes. Comprehensive experimental results validate our two contributions. We hope this work could benefit future works in this field.

External IDs:dblp:journals/ijon/LiangDCWC25