APTO：通过位置感知修剪加速基于序列化的点云变换器.pdf

《APTO：通过位置感知修剪加速基于序列化的点云变换器.pdf》由会员分享，可在线阅读，更多相关《APTO：通过位置感知修剪加速基于序列化的点云变换器.pdf（29页珍藏版）》请在三个皮匠报告上搜索。

1、30th Asia and South Pacific Design Automation Conference(ASPDAC 25)Tokyo,JapanJan.21,2025APTO:Accelerating Serialization-Based Point Cloud Transformers with Position-Aware PruningQichu Sun,Rui Meng,Haishuang Fan,Fangqiang Ding,Linxi Lu,Jingya Wu,Xiaowei Li,Guihai Yan1.State Key Laboratory of Process

2、ors,Institute of Computing Technology,Chinese Academy of Sciences 2.University of Chinese Academy of Sciences 3.University of Edinburgh 4.YUSUR Technology Co.,LPoint Cloud ProcessingPoint cloud is a representation of 3D data,containing valuable geometry&color information Autonomous Driving Robotic P

3、erceptionAR/VRAccurate,real-time and energy-efficient point cloud processing is crucialPoint Cloud TransformersPoint-Based ModelsSerialization-Based ModelsRepeated point access leads to redundant computation&memory use,while window size restricts accuracyRegular memory access,less redundant computat

4、ion,and better accuracyFarthest Point Sampling(FPS)&k-Nearest Neighbors(kNN)for down-samplingAttention mechanisms in local windows forfeature computationOrganizes points onto a directed curve(e.g.z-curve)3D sparse convolution for down-samplingLarger attention windows based on the curveFPS&kNNAttnSpC

5、onvAttnMotivation:Performance Bottleneck AnalysisInference time breakdown:Octformer:46%SpConv&27%attentionPTv3:32%SpConv&50%attentionFailed to meet real-time requirementsSpConv&attention are two key bottlenecks of serialization-based modelsSerialization-based modelsreal-time:30ms1 Peng-Shuai Wang.20

6、23.Octformer:Octree-based transformers for 3d point clouds.ACM Transactions on Graphics(TOG)42,4(2023),111.2 Xiaoyang Wu,Li Jiang,Peng-Shuai Wang,Zhijian Liu,Xihui Liu,Yu Qiao,Wanli Ouyang,Tong He,and Hengshuang Zhao.2024.Point Transformer V3:Simpler Faster Stronger.In Proceedings of the IEEE/CVF Co