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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Complete Structure Guided Point Cloud Completion via Cluster- and Instance-Level Contrastive Learning
Authors: Yang Chen, Yirun Zhou, Weizhong Zhang, Cheng Jin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both synthetic and real-world datasets demonstrate that our approach significantly outperforms state-of-the-art self-supervised methods. |
| Researcher Affiliation | Academia | Yang Chen1, Yirun Zhou1, Weizhong Zhang2,3, Cheng Jin1,3 1College of Computer Science and Artificial Intelligence 2School of Data Science, Fudan University 3Shanghai Key Laboratory of Intelligent Information Processing EMAIL, EMAIL |
| Pseudocode | No | The paper describes its methodology through narrative text and illustrative figures (Figure 2: The Pipeline of CSG-PCC, Figure 3: Illustration of Feature Permutation Consistency Constraint (FPCC)), but does not include a dedicated section or block explicitly labeled as 'Pseudocode' or 'Algorithm' with structured steps. |
| Open Source Code | No | The source code will be made publicly available after we complete its refactoring and documentation. |
| Open Datasets | Yes | We conducted experiments on the synthetic datasets 3D-EPN [9] and PCN [38]. Both of them are derived from the Shape Net [2] dataset... Moreover, we evaluate our method on real-world dataset Scan Net [8]. |
| Dataset Splits | No | Both of them are derived from the Shape Net [2] dataset, with the former containing more data, e.g., the chair class has 40,000 pairs for training in 3D-EPN while 5,750 pairs in PCN. |
| Hardware Specification | Yes | The experiments were conducted on four NVIDIA Ge Force RTX 3090 GPUs with 24GB memory each. |
| Software Dependencies | No | We train a separate network for each class using the Adam W optimizer with a starting learning rate of 10 3 and a weight decay of 10 3 for 300 epochs. |
| Experiment Setup | Yes | For the loss function, we set the λcompletion, λcluster, λinstance to 1, 0.1, and 0.01, respectively. The FPCC is computed every 50 backpropagation steps with a weight of 0.1. Like P2C [7], we divide the incomplete point cloud into 64 patches, each containing 32 points. The num of shape prototypes |M| is set to 32 and the temperature scaling parameter t is set to 0.4. We train a separate network for each class using the Adam W optimizer with a starting learning rate of 10 3 and a weight decay of 10 3 for 300 epochs. |