Action Recognition

最近关注了行为识别的领域，这个领域主要任务是视频分类，输入一个短视频，经过训练出来的分类器，得到正确的类别。但是一个视频如果存在多个行为，那么这个任务将会变成，输入一个短视频，经过预先训练的分类器，不仅要得到每一个行为的类别，还要得到行为开始时间和结束时间。这样，这个任务的难度提高不少，也更具有挑战性。
为了方便大家进行对比实验，我把目前的方法以及效果粘贴出来，持续更新。

Related Database

常见的数据集如下：
UCF101: http://crcv.ucf.edu/data/UCF101.php
HMDB51: http://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/
THUMOS 15:http://www.thumos.info/home.html
Sports-1M:http://cs.stanford.edu/people/karpathy/deepvideo/
YouTube-8M:https://research.google.com/youtube8m/download.html
ActivityNet:http://activity-net.org/download.html
FCVID：http://bigvid.fudan.edu.cn/FCVID/ (特别感谢姜育刚老师对FCVID的贡献)

Recent popular method’s results

Method HMDB51 UCF101 improved IDT 57.2% 85.9% IDT high-dim encode 61.1% 87.9% Two-stream[1] 59.4% 88% C3D + iDT + linear SVM[2] - 90.4% DOVF + MIFS [3] 75% 95.3% Very deep Two-stream[4] - 91.4% TSN [5] 69.4% 94.2% Multi-stream[6] - 92.2%

Reference:

[1]Two-Stream Convolutional Networks for Action Recognition in Videos

[2]Learning Spatiotemporal Features with 3D Convolutional Networks

[3]Deep local Video Feature for Action Recognition

[4]Towards Good Parctices for very deep two-stream convnets

[5]Temporal Segment Networks: Towards Good Practices for Deep Action Recognition 2016 ECCV

[6] Multi-Stream Multi-Class Fusion of Deep Networks for Video Classification