DeepDT: Learning Geometry From Delaunay Triangulation for Surface Reconstruction
Authors: Yiming Luo, Zhenxing Mi, Wenbing Tao2277-2285
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this part, we perform a series of experiments on datasets of different scales to qualitatively and quantitatively evaluate our Deep DT from different perspectives. Datasets & Evaluation Metrics We select three datasets with quite different types to comprehensively compare our Deep DT with traditional methods and other state-of-the-art learning-based methods. They are Shape Net (Chang et al. 2015), DTU (Jensen et al. 2014) and Stanford 3D. We choose the Chamfer-L1 distance and the normal consistency (NC) score for experiments on Shape Net. For DTU dataset, we follow the DTU Completeness metric given by DTU and also take Chamfer Distances (CD) into consideration. We compute the CD between the ground truth point cloud and the vertices of the result surface that is expressed by the triangular meshes. |
| Researcher Affiliation | Academia | National Key Laboratory of Science and Technology on Multi-spectral Information Processing School of Artifical Intelligence and Automation, Huazhong University of Science and Technology, China yiming luo@163.com, mizhenxing.henan@gmail.com, wenbingtao@hust.edu.cn |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a direct link to a code repository. |
| Open Datasets | Yes | Datasets & Evaluation Metrics We select three datasets with quite different types to comprehensively compare our Deep DT with traditional methods and other state-of-the-art learning-based methods. They are Shape Net (Chang et al. 2015), DTU (Jensen et al. 2014) and Stanford 3D. |
| Dataset Splits | Yes | For fair comparison, we adopt the same train/validation/test split (about 30K/4.5K/9K shapes) as the mentioned methods. |
| Hardware Specification | Yes | We test the efficiency of Deep DT on 1 Ge Force RTX 2080 Ti GPU in an Intel Xeon(R) CPU system with 40 2.20 GHz cores. |
| Software Dependencies | No | The paper mentions: "Delaunay triangulation is constructed from the point cloud using the Computational Geometry Algorithms Library (CGAL) (Boissonnat et al. 2000)." However, it does not specify a version number for CGAL or any other software components. |
| Experiment Setup | Yes | In the following experiments, unless otherwise specified, the number of reference locations (N ref) is set to 5, and the λ1, λ2 are set to 0.9 and 0.1, respectively. |