[20220606更新] 百度云分享功能好像又恢复了，链接：https://pan.baidu.com/s/1LZX-xBHqFbKGkqr-iBM2MA 提取码：stij。如果不行的话，可以添加我微信“smartporridge”，备注ava数据下载，点对点分享吧
[20200529更新] 百度云的好友点对点分享是可以的，可以添加我微信“smartporridge”，备注ava数据下载，点对点分享吧
[20200525更新] 分享链接被百度屏蔽了，说是侵权或色情，，，，需要点对点分享，有需要下载的可以添加我微信“smartporridge”，备注ava数据下载，点对点发链接吧。
！！更新百度云下载链接：链接: https://pan.baidu.com/s/1iUDhSMv9flaQo2ryjRmx7w 提取码: 4zxc
另外欢迎关注公众号“arXiv每日学术速递”，获取每日CV、NLP、AI等方向的学术信息哦,扫码关注
This repository contains the train and test annotations, all the videos' youtube IDs, action_id and some videos may be inaccessible on Youtube from your location.
仓库中有AVA的训练和测试annotations，也有所有视频的Youtube ID， 以及所有类别的labels，以及部分因为版权原因下载不到的视频下载方法。
The AVA dataset densely annotates 80 atomic visual actions in 57.6k movie clips with actions localized in space and time, resulting in 210k action labels with multiple labels per human occurring frequently. The main differences with existing video datasets are:
- the definition of atomic visual actions, which avoids collecting data for each and every complex action;
- precise spatio-temporal annotations with possibly multiple annotations for each human;
- the use of diverse, realistic video material (movies).
More details about the dataset and initial experiments can be found in this arXiv paper.
The AVA dataset contains 192 videos split into 154 training and 38 test videos. Each video has 15 minutes annotated in 3 second intervals, resulting in 300 annotated segments. These annotations are specified by two CSV files: ava_train_v1.0.csv and ava_test_v1.0.csv.
Each row contains an annotation for one person performing an action in an interval, where that annotation is associated with the middle frame. Different persons and multiple action labels are described in separate rows.
The format of a row is the following: video_id, middle_frame_timestamp, person_box, action_id
- video_id: YouTube identifier
- middle_frame_timestamp: in seconds from the start of the YouTube.
- person_box: top-left (x1, y1) and bottom-right (x2,y2) normalized with respect to frame size, where (0.0, 0.0) corresponds to the top left, and (1.0, 1.0) corresponds to bottom right.
- action_id: identifier of an action class, see ava_action_list_v1.0.pbtxt
The Youtube video IDs (train split) is ava_ytids_train_v1.0.txt The Youtube video IDs (test split) is ava_ytids_test_v1.0.txt
To download youtube videos, you need youtube-dl , for example:
youtube-dl -f best -f mp4 55Ihr6uVIDA -o "video_data/55Ihr6uVIDA.mp4"
Please note that some videos may be inaccessible on Youtube from your location.If you are a researcher interested in the whole dataset, click here for more information. And then you can download this part of videos after aome operations.
-FaXLcSFjUI.mp4 2017-07-17 14:50 365M
0f39OWEqJ24.mp4 2017-07-17 14:36 437M
2XeFK-DTSZk.mkv 2017-08-08 03:12 270M
4trIFq61-lk.mkv 2017-07-24 03:13 454M
EQZWzLyx-GM.mkv 2017-09-07 03:22 242M
F_-zE1dQsso.mkv 2017-08-02 03:24 225M
G3nRbyu0gMs.mp4 2017-07-17 14:51 218M
K--hW14uzA0.mkv 2017-09-12 03:29 877M
KHHgQ_Pe4cI.mkv 2017-08-08 03:29 845M
N0Dt9i9IUNg.mkv 2017-07-17 15:00 426M
PmElx9ZVByw.mp4 2017-07-17 15:04 431M
XIx-C22Ewk4.mp4 2017-08-09 03:47 788M
ZFQ3lF6yq_E.mkv 2017-08-01 03:49 459M
_2Isct32Msg.mkv 2017-08-26 03:11 627M
bnW1PXGt5hw.mp4 2017-07-17 14:46 379M
lSCEt_mCHlM.mkv 2017-07-17 14:59 2.6G
nxL0yqWP3H0.mkv 2017-07-17 15:01 659M
The dataset is made available by Google Inc. under a Creative Commons Attribution 4.0 International (CC BY 4.0) license.
-------华丽丽的分割线----------
谷歌发布了新的数据集 AVA（atomic visual actions），提供扩展视频序列中每个人的多个动作标签。AVA 包括 YouTube 公开视频的 URL，使用包含 80 个原子动作（atomic action）集进行标注（如「走路」、「踢（某物）」、「握手」），所有动作都有时空定位，从而产生 57.6k 视频片段、96k 标注人类动作和 210k 动作标签。
部分视频因为版权原因在国内下载不到，谷歌提供了高速缓存供大家下载，可以先在here注册成为学术用途，非常简单，不需要账号，同意不作商业用途之后，就可以得到账号和密码，既可以下载这部分视频。
