This repository contains code to download all subsets (700, 600, or 400) of Kinetics dataset. You can use the same download code to download 6 seconds long clips around the timestamps provided in Ava-Kinetics detection dataset. Code is modified from ActivityNet Crawler.
Orignal Crawler store videos in respective label directories. However, there is an overlap among the subsets and labels (class names) are changed, e.g. passing american football (not in-game) renamed to passing American football (not in-game) from 600 to 700, or class is made more fine-grained, e.g. picking fruit into multiple subclasses.
Which leads to multiple duplicate copies of some videos in different subfolder/directories.
Since we have labels in .csv files, then we do not need to store videos in respective label directories.
It provides two benefits, i) avoid same video being stored under different folder, i.e. no duplicate copies ii) directory structure is simple, i.e. all videos are stored under single directory.
First, clone this repository and make sure that all the submodules are also cloned properly.
git clone git@github.com:gurkirt/kinetics-download-prep.git
cd kinetics-download-prep
Next, setup your environment
conda install joblib
conda install pandas
conda install sqlite
conda install zlib
conda install -c menpo ffmpeg
pip install --upgrade youtube-dl
python based ffmpeg might not work. You can install system ffmpeg, which is usall install at /usr/bin/ffmpeg
Use reorg_kinetics_videos.py to reorgnise existing videos from videos_dir.
python reorg_kinetics_videos.py <videos_dir> <dst_dir>
If dst_dir doesn't exist the it will create it
Now, you can download the reminaing or all dataset splits by calling:
mkdir <output_dir>;
To download videos from specfic set.
python download.py <output_dir> --input_csv={dataset_split}.csv
OR, you can download entire dataset with all it's versions and all of it's sets. Just do not sepcify input_csv, by default it will pick up kinetics_csv directory and all the .csv files there.
Also, You can specify number of jobs to run in parllel by --n=number.
python download.py <output_dir>
python frame_extract.py <videos_dir> <output_dir> --fps=30
Extract frames at particular frame rate by --fps=int_number (default 30) or video fps via setting --fps=0. You can specify number of jobs to run in parllel by --n=number.
NOT WORKING at the momement, some problem with FFMPEG
Smallest side to 256. Skeleton is already there in downscale_videos.py.
python downscale_videos.py <input_dir> <output_dir> --fps=30
Also, You can specify number of jobs to run in parllel by --n=number.
NOTE AVA-Kinetics download as implemented here doesn't make any sense to me. -- Ross Wightman
Similiar to Kinetics you can download clip around the time stamps provide in AVA-Kinetics dataset.
First step to construction download list from AVA-kinetics annoatated videos and their time stamps. You can run make_ava_kin_download_csv.py for it. I have laready performed this task on AVA-kinetics-V.1. And you can find the download csv for AVA-Kinetics in ava_kinetics_csv.
Now, similiar to Kinetics, you can run download script.
python download.py <output_dir> --input_csv=ava_kinetics_csv\videos_to_download.csv
It will download all the clips with <videoname>_<timestamp>.mp4. Timestamp here is integer value, calculated from AVA-kinetics timestamps (let's call it ts), timestamp=math.floor(ts). Since, we have want a clip around that time stamp, I set the start-time of that clip to be 3 seconds before the timestamp and duration to be 6 seconds. To be precise, start-frame = max(0,timestamp-3). When you want to get annotated frame it can be obatined by frame_number = ts*fps if math.floor(ts)<=3 else (3 + ts-math.floor(ts))*fps of a video <videoname>_<timestamp>.mp4 extract at particular fps.
Use python-ffmpeg to load videos. Trial there in load_frame_eg.py, but it doesn't work that fast.
Try alternative ways to load video quick.
Extend it to video chunk loading not the whole video. Should be faster but no easy way to do it other than loading from frame dumps.