Geometry Sharing Network for 3D Point Cloud Classification and Segmentation
Authors: Mingye Xu, Zhipeng Zhou, Yu Qiao12500-12507
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on public datasets, Model Net40, Shape Net Part. Experiments demonstrate that GS-Net achieves the state-of-the-art performances on major datasets, 93.3% on Model Net40, and are more robust to geometric transformations. |
| Researcher Affiliation | Collaboration | Mingye Xu,1,2 Zhipeng Zhou,1 Yu Qiao1,3 1Shen Zhen Key Lab of Computer Vision and Pattern Recognition, SIAT-Sense Time Joint Lab,Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3SIAT Branch, Shenzhen Institute of Artificial Intelligence and Robotics for Society {my.xu, zp.zhou, yu.qiao}@siat.ac.cn |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described in this paper. No explicit code release statement or repository link is found. |
| Open Datasets | Yes | We conduct extensive experiments on public datasets, Model Net40, Shape Net Part. Model Net40 (Wu et al. 2015) contains 12,311 CAD models from 40 categories. Shape Net Part benchmark (Yi et al. 2016). |
| Dataset Splits | No | The paper specifies training and testing splits for ModelNet40 (9,843 models for training and 2,468 for testing) and ShapeNet Part (14,006 models for training and 2,874 for testing), but does not explicitly provide details for a separate validation split. |
| Hardware Specification | Yes | The forward time is recorded with a batch size of 8 on a single GTX 1080 GPU, which is the same hardware environment of the comparison models. |
| Software Dependencies | No | The paper mentions "These models are implemented by Pytorch." but does not provide specific version numbers for PyTorch or any other software dependencies needed to replicate the experiment. |
| Experiment Setup | Yes | As for classification task, instead of using only the last level s features as the encoder s output (Wang et al. 2018), we concatenate all levels features together and extract the holistic features by global max pooling and global average pooling. The concatenation of all levels features aims to fuse the features from different levels and the pooling operator urges to capture the most effective features for classification. Then we handle the holistic features by fully-connected layers with integrated dropout (Srivastava et al. 2014) to calculate the probability for each category. The cross-entropy loss is used for training. In the ablation study are conducted using k = 20 nearest neighbors. The forward time is recorded with a batch size of 8. We use FPS algorithm to down-sample the points and features at 3 levels (1024-512-256 points in classification network). |