Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Curvature Diversity-Driven Deformation and Domain Alignment for Point Cloud

Authors: Mengxi Wu, Hao Huang, Yi Fang, Mohammad Rostami

TMLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical experiment results show that our CDND achieves state-of-the-art performance by a noticeable margin over existing approaches. Our codes are available at: https://github.com/WMX567/CurvRec_DNWD. Extensive experiments on standard classification and segmentation benchmarks demonstrate that CDND achieves state-of-the-art performance compared to existing approaches. We evaluate our method on the Point DA-10 Qin et al. (2019) dataset, a domain adaptation dataset for point cloud classification, and on Point Seg DA Achituve et al. (2021), a dataset for point cloud segmentation. Ablation studies confirm that both components of CDND are necessary for optimal performance.
Researcher Affiliation Academia Mengxi Wu EMAIL Department of Computer Science University of Southern California Hao Huang EMAIL Department of Computer Science New York University Yi Fang EMAIL Department of Computer Science New York University Mohammad Rostami EMAIL Department of Computer Science University of Southern California
Pseudocode No The paper describes methods in text and equations, for example, Section 3 'Proposed Method' details the problem formulation, curvature diversity-driven deformation, and domain alignment. However, there are no explicitly labeled pseudocode or algorithm blocks, nor any figures or sections formatted as code-like steps for a procedure.
Open Source Code Yes Our codes are available at: https://github.com/WMX567/CurvRec_DNWD.
Open Datasets Yes We evaluate our method on the Point DA-10 Qin et al. (2019) dataset, a domain adaptation dataset for point cloud classification, and on Point Seg DA Achituve et al. (2021), a dataset for point cloud segmentation. For the Point DA-10 dataset, we compare our approach against the state-of-the-art methods for point cloud domain adaptation, including DANN Ganin et al. (2016), Point DAN Qin et al. (2019), RS Sauder & Sievers (2019), Def Rec+PCM Achituve et al. (2021), GAST Zou et al. (2021), Implicit PCDA Shen et al. (2022), and the most recent method with publicly available codes, PCFEA Wang & el al (2025). Point DA-10 consists of three domains: Shape Net-10 Chang et al. (2015), Model Net-10 Wu et al. (2015), and Scan Net-10 Dai et al. (2017), each sharing ten distinct classes. Point Seg DA consists of four domains: ADOBE, FAUST, MIT, and SCAPE.
Dataset Splits Yes We evaluate our method on the Point DA-10 Qin et al. (2019) dataset, a domain adaptation dataset for point cloud classification, and on Point Seg DA Achituve et al. (2021), a dataset for point cloud segmentation. The model is trained using source-labeled and target-unlabeled data. The specific hyperparameter values can be found in Table 5. Similarly, for the Point DA dataset, the hyperparameters are listed in Table 4. where α, γ, β1, β2 can be tuned using the target domain validation set and setting details can be found in Appendix A.2.
Hardware Specification Yes We trained our three CDND models with seeds {1, 2, 3} on A100 GPUs.
Software Dependencies No The paper mentions using DGCNN as the feature extractor and the Adam optimizer. It also states that the code is based on an open-source implementation. However, specific version numbers for these or other key software components (like Python, PyTorch, or CUDA) are not provided.
Experiment Setup Yes Hyperparameter settings and implementation details can be found in Appendix A.2. Hyperparameter Values Learning Rate 0.001, 0.0001 (S+M, MS) α 0.5 γ 0.5 β1 [0.0, 1.0] β2 0.2 # of Epochs 200. We use DGCNN as the feature extractor Achituve et al. (2021) for fair comparison. We repeat our experiments three times using distinct random seeds for initialization and report the average accuracy and standard deviation. To ensure a fair comparison, we maintain the same seed for data shuffling and use the Adam optimizer Kingma & Ba (2014) for optimization. For our method, we set the number of regions to be deformed to 5 for Point DA and 10 for Point Seg DA. Each region contains 55 points. The total number of regions is set to 20 for the Point DA dataset and 40 for the Point Seg DA dataset.