Modify Self-Attention via Skeleton Decomposition for Effective Point Cloud Transformer
Authors: Jiayi Han, Longbin Zeng, Liang Du, Xiaoqing Ye, Weiyang Ding, Jianfeng Feng808-816
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments have shown that our proposed SD-based transformer, is an effective approximation of the standard SA. We utilize the Simple Point Cloud Transformer (SPCT)(Guo et al. 2021) as our backbone network, and validate our approach on point cloud classification and segmentation tasks on the Model Net40 and Shape Net datasets, respectively. Extensive experiments demonstrate that SD-SA greatly improves the efficiency of SPCT, and achieves performance on both tasks comparable to that of the baseline model. |
| Researcher Affiliation | Collaboration | 1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China,... 2Interactive Entertainment Group, Tencent Inc., China. 3Baidu Inc., China. |
| Pseudocode | Yes | Algorithm 1: SD-SA |
| Open Source Code | No | The paper does not provide a specific link to the source code for their proposed method (SD-SA) nor does it explicitly state that their code is open-source. It references a third-party implementation ('Pointnet-Pointnet2-pytorch') used as a baseline, but this is not their own code. |
| Open Datasets | Yes | Dataset. We adopt the two most commonly used benchmark datasets to evaluate our SD-SA: Shape Net (Yi et al. 2016) and Model Net40 (Wu et al. 2015), respectively. We also validate our approach on 3d object detection on KITTI dataset (Geiger et al. 2013). |
| Dataset Splits | No | Shape Net contains 16880 3D models, 14006 of which are used for training, and 2874 are used for testing. Model Net40 contains 40 object categories and 12311 computer-aided design (CAD) models,the official division into 9843 training objects and 2468 testing objects is utilized. We also validate our approach on 3d object detection on KITTI dataset (Geiger et al. 2013). We utilize the official training and validation set in experiments. While train/test splits are provided for ShapeNet and ModelNet40, and a validation set is mentioned for KITTI, explicit numerical details for a validation split are not provided for all datasets, nor are all three splits (train/val/test) explicitly quantified for each dataset. |
| Hardware Specification | Yes | We use the Adam optimizer to train the model for 250 epochs on 3 NVIDIA GTX 1080 Ti (12 GB). |
| Software Dependencies | No | The paper references the 'Pointnet-Pointnet2-pytorch' framework, but does not specify exact version numbers for PyTorch or any other software dependencies needed for replication. |
| Experiment Setup | Yes | We use the Adam optimizer to train the model for 250 epochs on 3 NVIDIA GTX 1080 Ti (12 GB). For point cloud segmentation and object classification, the batch size is set to 128 and 64, respectively. A gradient clip is utilized, and the maximum norm is set to 10. |