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..
Coarse-to-Fine 3D Part Assembly via Semantic Super-Parts and Symmetry-Aware Pose Estimation
Authors: Xinyi Zhang, Bingyang Wei, Ruixuan Yu, Jian Sun
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the Part Net benchmark demonstrate that CFPA achieves state-of-the-art performance in assembly accuracy, structural consistency, and diversity across multiple categories. Experiments on Part Net show that CFPA outperforms prior methods in pose accuracy, structural consistency, and assembly diversity. We evaluate assembly accuracy on Chair, Table, and Lamp categories using SCD, PA, and CA, with PA and CA computed under a Chamfer distance threshold of 0.01. We also report PA and CA under varying thresholds (0.01-0.05), with performance curves in Figure 3 and average results in Table 2. We further evaluate assembly diversity using QDS and WQDS. As shown in Table 3, CFPA achieves the highest WQDS across all categories and the highest QDS on Chair, indicating its ability to generate shape diverse yet structurally valid assemblies. Table 4: Ablation study on super-parts. Table 5: Ablation study on designs in the coarse pose estimation stage and pose refinement stage. Table 6: Ablation study on symmetry-aware loss. Figure 4 presents the qualitative results on Chair, Table and Lamp categories. |
| Researcher Affiliation | Academia | 1School of Airspace Science and Engineering, Shandong University, Weihai 264209, China 2School of Mathematics and Statistics, Xi an Jiaotong University, Xi an 710049, China 3Pazhou Laboratory (Huangpu), Guangzhou, Guangdong 510555, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methods textually and visually through a pipeline figure (Figure 1), but it does not contain explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/ zhangxinyi364/CFPA. |
| Open Datasets | Yes | We evaluate our method on the Part Net dataset [56], focusing on the three largest object categories: Chair (6,323 shapes), Table (8,218 shapes), and Lamp (2,207 shapes). We follow the official data splits for every categories, using 70% of the shapes for training, 10% for validation, and the remaining 20% for testing. |
| Dataset Splits | Yes | We follow the official data splits for every categories, using 70% of the shapes for training, 10% for validation, and the remaining 20% for testing. |
| Hardware Specification | Yes | The model is trained for 500 epochs using a batch size of 64 across 4 NVIDIA RTX 4090 GPUs. |
| Software Dependencies | No | We implement CFPA in Py Torch [51] using Adam W optimizer [52] with batch size 64. While PyTorch is mentioned, a specific version number for PyTorch or other libraries is not provided. |
| Experiment Setup | Yes | We implement CFPA in Py Torch [51] using Adam W optimizer [52] with batch size 64. Following [6, 7, 15, 17], all the input parts are centralized and normalized by PCA and added with random noise during training. We adopt Min-of-N (Mo N) strategy [53] for optimization. The model learns M=16 super-parts and uses 8-head attention in all MHA layers. Key hyper-parameters are set as ϵ=10 3, γ=10, and λ=0.1. The model is trained for 500 epochs using a batch size of 64 across 4 NVIDIA RTX 4090 GPUs. We adopt the Adam W optimizer with an initial learning rate of 7.5 10 5 and a weight decay of 1 10 4. A cosine annealing learning rate schedule is applied with a decay factor of 100 to progressively reduce the learning rate over the training process. |