【論文筆記】2020CVPR - PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition

本文轉載自查看原文 2020-10-16 19:48 543 Action Recognition

任務

action recognition

對骨架節點坐標進行視角不變預處理

We align the action sequences by implementing a view-invariant transformation which transforms keypoints coordinates from original coordinate system into a view-invariant coordinate

system。

基本流程

先訓練出一個encoder-decoder，這個encoder是雙向GRU，decoder是單向GRU。把encoder得到的特征叫做high dimensional feature。如下圖：

再把high-dimensional feature送進另一個encoder-decoder，這里的encoder和decoder都是全連接層，激活函數是tanh，如下圖所示。把這個encoder得到的特征叫做low dimensional feature。用這個low dimensional feature做KNN聚類。

數據集

North-Western UCLA (NW-UCLA) dataset

UWA3D Multiview Activity II (UWA3D) dataset

NTU RGB+D dataset

免責聲明！

本站轉載的文章為個人學習借鑒使用，本站對版權不負任何法律責任。如果侵犯了您的隱私權益，請聯系本站郵箱yoyou2525@163.com刪除。

猜您在找 【論文翻譯】：2020-CVPR-PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition 【論文筆記】Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition 【論文筆記，action recognition,動作識別】論文筆記 | A Closer Look at Spatiotemporal Convolutions for Action Recognition Skeleton-Based Action Recognition with Directed Graph Neural Network 論文筆記：Unsupervised Domain Adaptation by Backpropagation Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition (ST-GCN) 論文筆記【MDNet//CVPR2016】 Two-Stream Adaptive Graph Convolutional Network for Skeleton-Based Action Recognition GHM論文筆記（CVPR2019）