Introduction to Video Action Recognition

1. Classification of Video Timing Action Recognition Algorithms

According to the working mode of the network, video timing action recognition algorithms can be roughly divided into four categories:

• Method using 2D convolution
• Using the method of 3D convolution
• double flow method
• The method of introducing VLAD

1.1 Method using 2D convolution

• "TSM: Temporal Shift Module for Efficient Video Understanding" algorithm detailed explanation
• "TEA: Temporal Excitation and Aggregation for Action Recognition" algorithm detailed explanation
• "TDN: Temporal Difference Networks for Efficient Action Recognition" paper detailed explanation
• "No frame left behind: Full Video Action Recognition" algorithm details

1.2 Method using 3D convolution

• "Learning Spatiotemporal Features with 3D Convolutional Networks" (C3D) algorithm details
• 《Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification》(S3D)算法详解
• Detailed Algorithm Explanation of "ECO: Efficient Convolutional Network for Online Video Understanding"
• "Learning Spatio-Temporal Representation with Pseudo-3D Residual Networks" (P3D) Algorithm Details
• Detailed explanation of "SlowFast Networks for Video Recognition" algorithm
• "X3D: Expanding Architectures for Efficient Video Recognition" algorithm detailed explanation

1.2 Dual flow method

• Detailed explanation of "Two-Stream Convolutional Networks for Action Recognition in Videos" (TSN) algorithm
• Detailed explanation of "Temporal Segment Networks: Towards Good Practices for Deep Action Recognition" (TSM) algorithm

1.3 The method of introducing VLAD

• "ActionVLAD Algorithm Details"

2. Introduction to common data sets

Sports-1M数据集介绍：
	* 1.1 millions运动视频
	* 487个视频类

UCF101数据集介绍：
	* 13320个视频片段
	* 9.5K训练，3.7K测试视频
	* 视频帧大小320*240
	* 总共101类，内容包含化妆刷牙、爬行、理发、弹奏乐器、体育运动五大类。
	* 每类动作由25个人做动作，每人做4-7组

ActivatyNet数据库介绍

* 人类动作识别数据库
* v1.3版本中有19994段视频，包含200类
* 10024段视频为训练集，4926段视频为验证集，5044段视频为测试集
* 测试集label没有公开，一般就是使用验证集来作为测试集

HMDB51数据介绍
	* 6766个视频
	* 51个动作类别
	* 内容包括人面部、肢体、和物体交互的动作这几大类

Kinetic-400 数据库介绍

* 240k训练视频，20k验证，35k测试
* 400类人类动作类别
* 内容为画画、大笑、拥抱、除草等
* 每个视频大约10秒
* 数据来源于YouTube

Kinetic-600 数据库介绍

* Kinetic-400数据库的扩展
* 600类人类动作类别
* 总共500k段视频

Charades 数据库介绍

* 9848段视频
* 157类室内日常行为
* 多标签
* 每个视频大约30s

For other video task introductions, please check-article <<Introduction to Mainstream Video Action Algorithm Tasks>>