Multi-Shape Matching with Cycle Consistency Basis via Functional Maps
Authors: Yifan Xia, Tianwei Ye, Huabing Zhou, Zhongyuan Wang, Jiayi Ma
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several public datasets demonstrate the superiority of our approach over current state-of-the-art methods. |
| Researcher Affiliation | Academia | 1Electronic Information School, Wuhan University, Wuhan 430072, China 2School of Computer Science and Engineering, Wuhan Institute of Technology, Wuhan 430205, China 3School of Computer Science, Wuhan University, Wuhan 430072, China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Pairwise Correspondences Refinement Algorithm 2: Overall Process of Our Approach |
| Open Source Code | Yes | Code https://github.com/Ye Tianwei/Cy Co Match |
| Open Datasets | Yes | FAUST (Bogo et al. 2014) contains 100 shapes representing 10 poses from 10 different human subjects, with each shape containing 6,890 vertices. SCAPE (Anguelov et al. 2005) includes 71 shapes, each depicting a different pose of the same human subject, with each shape comprising 12,500 vertices. TOSCA (Bronstein, Bronstein, and Kimmel 2008) consists of 76 shapes across 8 categories, including both animal and human forms. |
| Dataset Splits | Yes | FAUST (Bogo et al. 2014) contains 100 shapes representing 10 poses from 10 different human subjects, with each shape containing 6,890 vertices. For quantitative evaluation, we group the 10 poses of each subject into 10 independent shape collections. SCAPE (Anguelov et al. 2005) includes 71 shapes, each depicting a different pose of the same human subject, with each shape comprising 12,500 vertices. For quantitative analysis, we randomly select 10 shapes to create a shape set, resulting in a total of 7 shape sets. TOSCA (Bronstein, Bronstein, and Kimmel 2008) consists of 76 shapes across 8 categories, including both animal and human forms. Each shape contains approximately 10,000 vertices and includes ground truth data for benchmarking. For quantitative evaluation, we treat each category as an individual shape set. |
| Hardware Specification | Yes | Experiments are conducted on a desktop computer with a 3.50GHz Intel Core i9-9920X CPU and MATLAB R2018a, with GPU-accelerated K-nearest neighbor searches. |
| Software Dependencies | Yes | Experiments are conducted on a desktop computer with a 3.50GHz Intel Core i9-9920X CPU and MATLAB R2018a, with GPU-accelerated K-nearest neighbor searches. |
| Experiment Setup | Yes | Parameter α balances the contributions of Cl ij and Cr k in the final Cij from Eq. (8) and varies from 0.1 to 0.9. We use the TOSCA (Bronstein, Bronstein, and Kimmel 2008) dataset as a test set to determine the appropriate value. As shown in Fig. 4, both relative geodesic error and cycle error are minimized when α is set to 0.3. Parameter {ϵit}Max It it=1 is the key threshold for determining landmarks at each iteration, with its length corresponding to the value of Max It. Empirically, We set {ϵit}Max It it=1 = [0.12, 0.1] and Max It = 2. This configuration balances geodesic error, cycle error, and runtime based on the Local Mapping Distortion values during actual runs. We empirically set the parameters ρ = 10−3, β = µ = 1, and run ADMM until convergence. Competitor methods use author-provided settings and code, with the number of eigenfunctions set to 500 or the maximum dimension of Zoomout and Consistent Zoomout upsampling iterations. |