跳转至

行为识别-数据集&工具#

1.数据集-动态视频#

数据 基本情况 链接
HMDB-51 51个类别,6766个视频 https://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/#dataset
UCF-101 101类别,13320个短视频 https://www.crcv.ucf.edu/research/data-sets/ucf101/
Kinetics-700 650000个,用于ActivityNet比赛 https://deepmind.com/research/open-source/open-source-datasets/kinetics/
ava google。多人不同动作数据集 https://research.google.com/ava/explore.html
ActivityNet

HMDB-51: 压缩包大概6G,主要包含类别
- 常见面部动作(smile,laugh,chew,talk)
- 复杂面部动作
- 直梯动作
- 多人交互

-w971
-w1001

UCF-101: 包含的数据类型
- 人物交互
- 人体动作
- 人人交互
- 乐器演奏
- 体育运动

Kinetics
- 单人行为,例如绘画,喝酒,大笑,抽拳;
- 人人行为,例如拥抱,亲吻,握手;
- 人物行为,例如打开礼物,修剪草坪,洗碗。

ava-action:
包含了80个原子动作(比如走路,握手。。) 数据包含三部分:
- person pose:
- 人-物交互
- 人人交互
https://research.google.com/ava/explore.html
一共430个video,其中235个train, 64个validation,131 test。 每个视频是15min,每1s钟进行以此标注。
有一个机遇faster-rcnn的预训练模型。 tensorflow的目标检测api https://github.com/tensorflow/models/tree/master/research/object_detection#tensorflow-object-detection-api

数据格式 ava_train_v2.2.csv
video_id, middle_frame_timestamp, person_box, action_id, person_id

stand, sit, watch some body, listen to sb....

toread
https://www.jianshu.com/p/a4cc71126796

2.数据集-静态图像#

https://blog.csdn.net/lgk1996/article/details/79814888

data basic description link
Stanford40 Dataset 斯坦福-李飞飞团队穿件,包括40中不同的行为类别,9532张图片 http://vision.stanford.edu/Datasets/40actions.html
MPII Pose Dataset 主要是human pose estimation

Stanford40 Dataset: runing, phoning, play,drinking等, taking photos
B. Yao, X. Jiang, A. Khosla, A.L. Lin, L.J. Guibas, and L. Fei-Fei. Human Action Recognition by Learning Bases of Action Attributes and Parts. Internation Conference on Computer Vision (ICCV), Barcelona, Spain. November 6-13, 2011.

MPII Pose Dataset

3.工具#

常用的工具:
https://zhuanlan.zhihu.com/p/38597956

tool 说明 语言 链接
openpose 多人,2D,实时 python-API https://github.com/CMU-Perceptual-Computing-Lab/openpose
facebook-densePose mask-RCNN的变体 github.com/facebookresearch/Densepose
alphaPose tf, pytorch
Human Body Pose Estimation
Deeppose(2014)

(1)Open-mmlab家族#

MMSkeleton
原理的名字叫 st-gcn,后来改为 MMSkeleton
st-gcn: 比较老了 https://hub.docker.com/r/jaehwankimneo/st-gcn
mmskelton: https://hub.docker.com/r/ioir123ju/mmskeleton/tags

python setup.py develop

中间可能会报很多错,需要一步步的改
环境配置http://blog.sina.com.cn/s/blog_679f93560102wpyf.html

安装好mmskelton后,可以选安装mmdetection,这个参照其github上的说明来就行

mmcv

mmcv是一个MMLAB开发的cv的基础工具库,用来支持mmdetection,mmaction等

(2)openpose#

运行方式:
1. 命令行: ./build/examples/openpose/openpose.bin xxx
2. c++接口
3. Python接口

docker: docker pull exsidius/openpose

还有在google colab上进行尝试https://github.com/CMU-Perceptual-Computing-Lab/openpose/issues/949#issue-387855863

! apt update
! apt install -y cmake sudo libopencv-dev
! git clone https://github.com/CMU-Perceptual-Computing-Lab/openpose.git
! cd openpose/ubuntu && ./install_cmake.sh && ./install_cuda.sh && ./install_cudnn.sh
! cd openpose && git pull origin master && rm -r build || true && mkdir build && cd build && cmake .. && make -j`nproc`

# example demo usage
!cd openpose && ./build/examples/openpose/openpose.bin --video examples/media/video.avi --write_json output/ --display 0 --render_pose 0

cpu-only的版本:要求8G内存

安装

git clone https://github.com/CMU-Perceptual-Computing-Lab/openpose
cd build/
make -j`nproc`

(3) gluon#

https://gluon-cv.mxnet.io/model_zoo/action_recognition.html