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].
Point Cloud Mamba: Point Cloud Learning via State Space Model
Authors: Tao Zhang, Haobo Yuan, Lu Qi, Jiangning Zhang, Qianyu Zhou, Shunping Ji, Shuicheng Yan, Xiangtai Li
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Point Cloud Mamba surpasses the state-of-the-art (SOTA) point-based method Point Ne Xt and achieves new SOTA performance on the Scan Object NN, Model Net40, Shape Net Part, and S3DIS datasets. It is worth mentioning that when using a more powerful local feature extraction module, our PCM achieves 79.6 m Io U on S3DIS, significantly surpassing the previous SOTA models, De LA and PTv3, by 5.5 m Io U and 4.9 m Io U, respectively. Experiments We conduct experiments on four datasets: Scan Object NN (Uy et al. 2019) and Model Net40 (Wu et al. 2015) classification datasets, Shape Net Part (Chang et al. 2015) part segmentation dataset, and S3DIS (Armeni et al. 2016) semantic segmentation dataset. Ablation Analysis and Visualization Effect of serialization strategies. The key to applying Mamba for point cloud modeling is transforming point clouds into point sequences. As shown in Table 4, we conduct the ablation experiment with different serialization strategies. |
| Researcher Affiliation | Collaboration | Tao Zhang1,2 *, Haobo Yuan1, Lu Qi1, Jiangning Zhang4, Qianyu Zhou5, Shunping Ji1 , Shuicheng Yan2,3, Xiangtai Li2,3 1 Wuhan University, China 2 Skywork AI, Singapore 3 Nanyang Technological University, Singapore 4 Youtu Lab, Tencent, China 5 Shanghai Jiao Tong University, China |
| Pseudocode | No | The paper describes methods through textual descriptions and mathematical formulations (e.g., equations 1-7) but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/zhang-tao-whu/PCM |
| Open Datasets | Yes | Experiments We conduct experiments on four datasets: Scan Object NN (Uy et al. 2019) and Model Net40 (Wu et al. 2015) classification datasets, Shape Net Part (Chang et al. 2015) part segmentation dataset, and S3DIS (Armeni et al. 2016) semantic segmentation dataset. |
| Dataset Splits | Yes | Following Point MLP (Ma et al. 2022) and Point Ne Xt (Qian et al. 2022), we conducted experiments on PB_T50_RS, the most challenging and commonly used Scan Object NN variant. Averaged results in three random runs using 1024 points as input without voting are reported. For the detailed experiment settings, please refer to the supplementary materials. |
| Hardware Specification | No | The paper does not provide specific hardware details (like GPU/CPU models, memory, or specific cloud instances) used for running the experiments in the main text. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., Python, PyTorch, CUDA versions) needed to replicate the experiment in the main text. |
| Experiment Setup | Yes | Averaged results in three random runs using 1024 points as input without voting are reported. For the detailed experiment settings, please refer to the supplementary materials. Architecture PCM-Tiny PCM Mamba layers {1, 1, 2, 2} {1, 2, 2, 4} Serialization {[xyz]-[xzy]-[yxz, yzx]-[zxy-zyx]} {[xyz]-[xzy, yxz]-[yzx, zxy]-[zyx, H, z, z-trans]} Channels {192, 192, 384, 384} {384, 384, 768, 768} Order Prompts 6 6 Table 1: Architecture settings. For better extraction of local point features, we cascade 4 De LA blocks before PCM as an additional local feature extractor; please refer to the supplementary for details. |