【计算机视觉】行为识别(action recognition)相关资料-白红宇

【计算机视觉】行为识别(action recognition)相关资料

阅读量：4956 次

发布时间：2019-06-12

本文共 11247 字，大约阅读时间需要 37 分钟。

================华丽分割线=================这部分来自知乎====================

链接：http://www.zhihu.com/question/33272629/answer/60279003

有关action recognition in videos, 最近自己也在搞这方面的东西，该领域水很深，不过其实主流就那几招，我就班门弄斧说下video里主流的：

Deep Learning之前最work的是INRIA组的Improved Dense Trajectories(IDT) + fisher vector, paper and code:

基本上INRIA的东西都挺work 恩..

然后Deep Learning比较有代表性的就是VGG组的2-stream:

其实效果和IDT并没有太大区别，里面的结果被很多人吐槽难复现，我自己也试了一段时间才有个差不多的数字。

然后就是在这两个work上面就有很多改进的方法，目前的state-of-the-art也是很直观可以想到的是xiaoou组的IDT+2-stream:

还有前段时间很火，现在仍然很多人关注的G社的LSTM+2-stream:

然后安利下zhongwen同学的paper:

最后你会发现paper都必需和IDT比，

================华丽分割线=================这部分也来自知乎====================

链接：http://www.zhihu.com/question/33272629/answer/60163859

视频方面的不了解，可以聊一聊静态图像下的~

[1] Action Recognition from a Distributed Representation of Pose and Appearance, CVPR,2010

[2] Combining Randomization and Discrimination for Fine-Grained Image Categorization, CVPR,2011

[3] Object and Action Classification with Latent Variables, BMVC, 2011

[4] Human Action Recognition by Learning Bases of Action Attributes and Parts, ICCV, 2011

[5] Learning person-object interactions for action recognition in still images, NIPS, 2011

[6] Weakly Supervised Learning of Interactions between Humans and Objects, PAMI, 2012

[7] Discriminative Spatial Saliency for Image Classification, CVPR, 2012

[8] Expanded Parts Model for Human Attribute and Action Recognition in Still Images, CVPR, 2013

[9] Coloring Action Recognition in Still Images, IJCV, 2013

[10] Semantic Pyramids for Gender and Action Recognition, TIP, 2014

[11] Actions and Attributes from Wholes and Parts, arXiv, 2015

[12] Contextual Action Recognition with R*CNN, arXiv, 2015

[13] Recognizing Actions Through Action-Specific Person Detection, TIP, 2015

2010之前的都没看过，在10年左右的这几年（11,12）主要的思路有3种：1.以所交互的物体为线索（person-object interaction），建立交互关系，如文献5,6；2.建立关于姿态（pose）的模型，通过统计姿态（或者更广泛的，部件）的分布来进行分类，如文献1,4，还有个poselet上面好像没列出来，那个用的还比较多；3.寻找具有鉴别力的区域（discriminative），抑制那些meaningless 的区域，如文献2,7。10和11也用到了这种思想。

文献9,10都利用了SIFT以外的一种特征：color name，并且描述了在动作分类中如何融合多种不同的特征。

文献12探讨如何结合上下文（因为在动作分类中会给出人的bounding box）。

比较新的工作都用CNN特征替换了SIFT特征（文献11,12,13），结果上来说12是最新的。

静态图像中以分类为主，检测的工作出现的不是很多，文献4,13中都有关于检测的工作。可能在2015之前分类的结果还不够promising。现在PASCAL VOC 2012上分类mAP已经到了89%，以后的注意力可能会更多地转向检测。

================华丽分割线=================这部分来自互联网====================

[1] (干货较多，可以进去浏览浏览)

[2]

Tian, YingLi, et al. "Hierarchical filtered motion for action recognition in crowded videos." Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on 42.3 (2012): 313-323.
1. A new 3D interest point detector, based on 2D Harris and Motion History Image (MHI). Essentially, 2D Harris points with recent motion are selected as interest point.
2. A new descriptors based on HOG on image intensity and MHI. Some filtering is performed to remove cluttered motion and normalize descriptors.
3. KTH and MSR Action dataset

Yuan, Junsong, Zicheng Liu, and Ying Wu. "Discriminative subvolume search for efficient action detection." Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009.
1. A discriminative matching techniques based on mutual information and nearest neighbor algorithm
2. A better upper bound for Branching and Bounding to locate matched action that maximize mutual information
3. The key idea is to decompose the search space into spatial and temporal.

Lampert, Christoph H., Matthew B. Blaschko, and Thomas Hofmann. "Beyond sliding windows: Object localization by efficient subwindow search." Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on. IEEE, 2008.
1. Code online: (Efficient Subwindow Search)
2. Reducing the complexity of sliding window from n4 to averagely n2
3. Branching and Bounding techniques
4. Relies on a bounding funtion that gives a upper bound of the scoring function over a set of potential box
5. works well with linear classifiers and BOW features.

Li, Li-Jia, et al. "Object Bank: A High-Level Image Representation for Scene Classification & Semantic Feature Sparsification." NIPS. Vol. 2. No. 3. 2010.
1. Images are represented as a scale-invariant map of object detector response
2. Detectors are applied to novel images in multiple scales. At each scale, a 3 level spatial pyramid is applied. Responses are concatenated to form the descriptors for the image.
3. 200 objecst are selected from a 1000 objects pool
4. Evaluated In Scene classification task
5. L1 and L1/L2 regularized LR is applied to discover sparsity. The the L1/L2 group sparsity, group is defined for each object, hence object level sparsity. Bear in mind that there are multiple entries in the descriptors for each object. (marginal improvements)

Wang, Heng, et al. "Dense trajectories and motion boundary descriptors for action recognition." International journal of computer vision 103.1 (2013): 60-79.
1. Tracking over densely sampled points to get trajectories, in contrast with local representation. Not really dense sampling, grids are filtered by minEigen value criterion (Shi and Tomasi)
2. Motion boundary (derivative over optical flow field), to overcome camera motion
3. Code online:
4. Optical Flow field is filtered by Median Filter. based on opencv
5. Limit trajectory to overcome drift. Filter static point and error trajectories.
6. Trajectory shape, HOG, HOF and MBH descriptors along the trajectory
7. KTH (94.2%), Youtube (84.1%), Hollywood2 (58.2%), UCF Sports (88.0%), IXMAS (93.5%), UIUC (98.4%), Olympic Sports (74.1%), UCF50 (84.5%), HMDB51 (46.6%)

Liang, Xiaodan, Liang Lin, and Liangliang Cao. "Learning latent spatio-temporal compositional model for human action recognition." Proceedings of the 21st ACM international conference on Multimedia. ACM, 2013.
1. Laptev STIP with HOF and HOG, with BOW quantization
2. Leaf node for detecting action parts
3. Or node to account for intra-class variability
4. And node to aggregate action in a frame
5. Root node to identify temporal composition
6. Contextual interaction (connecting leaf nodes)
7. Everything is formulated in a latent SVM framework and solved by CCCP
8. Since the leaf node can move around from one Or-node to another, a reconfiguration step is used to rearrange the feature vector
9. UCF Youtube and Olympic Sports dataset

Sadanand, Sreemanananth, and Jason J. Corso. "Action bank: A high-level representation of activity in video." Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on. IEEE, 2012.
1. 98.2% KTH, 95.0% UCF Sports, 57.9% UCF50, 26.9% HMDB51
2. 205 video clips used as template to detect action from novel video.
3. Detectors are sampled from multi viewpoint and run with multiple scales
4. Output of detectors are maxpooled for ST volume through various pooling unit
5. "Action Spoting" for template detector
6. Code online:

Liu, Jingen, Benjamin Kuipers, and Silvio Savarese. "Recognizing human actions by attributes." Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on. IEEE, 2011.
1. 22 manually selected action attributes as semantic representation
2. Data Driven attributes as complementary information
3. Attributes as latent variable, just the parts in DPM model
4. Account for the class matching, attribute matching, attributes cooccurcance.
5. STIP by 1D-Gabor detector. Gradient based + BOW over ST volume
6. UIUC dataset, KTH, Olympic Sports Dataset

Niebles, Juan Carlos, Hongcheng Wang, and Li Fei-Fei. "Unsupervised learning of human action categories using spatial-temporal words." International Journal of Computer Vision 79.3 (2008): 299-318.
1. Unsupervised video categorizaton, using pLSA and LDA
2. Action Localization
3. Laptev's STIP is too sparse comparing with Dollar's
4. Simple gradient based descriptors and PCA applied to reduce dimensionality --> rely on codebook to deal with invariance
5. K-means with Euclidean distance metric
6. pLSA or LDA on top of BOW (# topic is equal to the categories to be recognized)
7. Each STIP is associated with a BOW, hence topic distribution, so it's trivial to perform Localization

Laptev, Ivan, et al. "Learning realistic human actions from movies." Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on. IEEE, 2008.
1. Annotating videos by aligning transcriptes
2. A movie dataset
3. Space-Time interest points + HOG + HOF around a ST volume
4. ST BOW. Given a video sequence, multiple way to segment it, each of which is called a channel
5. Multi-Channel \chi^2 kernel classification. Channel selection using greedy shrink
6. KTH (91.8%) and Movie (18.2% ~ 53.3%) dataset
7. STIP + HOG and HOF code:

[3]

Links to Datasets:

D. Weinland, R. Ronfard, E. Boyer

M. Blank, L. Gorelick, E. Shechtman, M. Irani, R. Basri

C. Schuldt, I. Laptev and B. Caputo

Y. Shi, Y. Huang, D. Minnen, A. Bobick, I. Essa

Y. Huang, I. Essa

Recent Action Recognition Papers:

D. Weinland, R. Ronfard, E. Boyer (CVIU Nov./Dec. '06)
11 actors each performing 3 times 13 actions: Check Watch, Cross Arms, Scratch Head, Sit Down, Get Up, Turn Around, Walk, Wave, Punch, Kick, Point, Pick Up, Throw.
Multiple views of 5 synchronized and calibrated cameras are provided.

A. Yilmaz, M. Shah (ICCV '05)
18 Sequences, 8 Actions: 3 x Running, 3 x Bicycling, 3 x Sitting-down, 2 x Walking, 2 x Picking-up, 1 x Waving Hands, 1 x Forehand Stroke, 1 x Backhand Stroke

Y. Sheikh, M. Shah (ICCV '05)
6 Actions: Sitting, Standing, Falling, Walking, Dancing, Running

M. Blank, L. Gorelick, E. Shechtman, M. Irani, R. Basri (ICCV '05)
81 Sequences, 9 Actions, 9 People: Running, Walking, Bending, Jumping-Jack, Jumping-Forward-On-Two-Legs, Jumping-In-Place-On-Two-Legs, Galloping-Sideways, Waving-Two-Hands, Waving-One-Hand Ballet

A. Yilmaz, M. Shah (CVPR '05)
28 Sequences, 12 Actions: 7 x Walking, 4 x Aerobics, 2 x Dancing, 2 x Sit-down, 2 x Stand-up, 2 x Kicking, 2 x Surrender, 2 x Hands-down, 2 x Tennis, 1 x Falling