| 摘 要: 三维点云识别在自动驾驶、虚拟现实等领域有着广泛的应用,这些领域对实时性要求较高。针对这一问题提出一种利用空间稀疏性来加速点云模型的方法。点云模型的正确率取决于少部分特征,通过设计一个轻量级模块估计给定当前特征的重要性,裁剪对结果影响较小的令牌(token)来减少模型计算量,将该模块添加到不同的层。在点云数据集 ModelNet40进行点云识别任务。通过修剪66%的输入token,相比于原模型减少了31%~35%的FLOPs,性能降低在0.5%以下。实验结果验证了该方法的有效性和可行性。 |
| 关键词: 点云识别 注意力模型 模型轻量化 |
|
中图分类号: TP391
文献标识码: A
|
|
| Acceleration Method for Point Cloud Attention Model Based on Dynamic Sparsification |
|
LIU Zilong1, BAO Yanxia2, XIAO Guolei3, SHEN Yang1,2,LAN Xiayan4
|
(1.School of Information Science and Engineering, Zhejiang Sc-i Tech University, Hangzhou 310018, China;
2.School of Mathematics and Computer, Lishui University, Lishui 323000, China;
3.State Grid Zhejiang Electric Power Co., Ltd.Yunhe County Power Supply Company, Lishui 323000, China;
4.Zhejiang Legu Technology Co., Ltd., Lishui 323000, China)
1906789057@qq.com; tlsheny@126.com; 13735979916@163.com; tlsheny@163.com; 365046719@qq.com
|
| Abstract: 3D point cloud recognition has extensive applications in fields such as autonomous driving and virtual reality, which demand high rea-l time performance. To address this issue, a method leveraging spatial sparsity to accelerate point cloud models is proposed. The accuracy of point cloud models depends on a small subset of features.By designing a lightweight module to estimate the importance of given current features and pruning tokens with minimal impact on the results, the computational load of the model is reduced. This module is incorporated into
different layers. Point cloud recognition tasks were conducted on the ModelNet40 dataset. By pruning 66% of the input tokens, the method achieved a reduction of 31% to 35% in FLOPs compared to the original model, with a performa |
| Keywords: point cloud recognition attention model model lightweighting |
