Video annotation tool for deep learning training labels.
This project enables users to generate Ground-Truth to train deep learning models.
Features supported: 1) Easy API Platform to load extant deep learning models. 2) UI to trigger inference and generate the results in a JSON format for the ground-truth. 3) human interface to correct the above ground truth collected.
This way we can reduce the human effort required to manually generate the ground-truth.
README From BeaverDam which we forked to implement automatic annotation:
This tool is for drawing object bounding boxes in videos. It also includes support for Amazon Mechanical Turk. See the paper.
With small amount of changes, you can also:
This tool currently does not support semantic segmentation.
cd BeaverDambrew install python3 (Mac) or sudo apt-get install python3 (Ubuntu)pip3 install virtualenv or maybe sudo pip3 install virtualenvscripts/setupscripts/seed -fWhen running any ./manage.py commands, use source venv/bin/activate to enter venv first.
See /deployment for tips on using BeaverDam for production.
Replace the credentials below with your own:
export AWS_ID="AKIAAAAYOURIDHERE"
export AWS_KEY="YOURmturkKEYhere5DyUrkm/81SRSMG+5174"When ready for real turkers, edit MTURK_SANDBOX to False in settings.py.
It is recommended to use IAM keys with only mturk permissions instead of root key.
scripts/serveThen navigate to localhost:5000 in your browser.
Need to run on a custom port? env PORT=1234 scripts/serve
For actual production deployment, we recommend using standard Django deployment procedures. Sample scripts using uWSGI & nginx are provided in /deployment. Remember to set DEBUG=False in settings.py.
Login is required to authenticate any changes. Turkers do not require accounts and are authenticated by BeaverDam via Mechanical Turk.
To make a superuser account, run inside venv ./manage.py createsuperuser
If you are using sample data, login with username test and password password.
Additional non-turker worker accounts can be created via /admin.
To add videos via web UI, navigate to /admin and create Video objects.
Alternatively, use ./manage.py shell, and create annotator.Video objects and call video.save().
Helper methods exist to create large number of video objects at once, see annotator/models.py.
Video objects can either be H.264 encoded video (See scripts/convert-to-h264), or a list of frames provided in the attribute image_list.
By using single-frame videos, BeaverDam can be used for image annotation.
Video annotations can be accessed via admin, /annotation/video_id, or through the Video objects' annotation attribute through the shell.
Tasks are created in the same way as Videos.
Only the video attribute needs to be filled out at creation time.
They can be published to mturk by calling task.publish().
To see what video pages look like on mturk preview mode, set url param preview=true.
For mturk's HIT accepted mode, set url param mturk=true.
Example: localhost:5000/video/0/?mturk=true
Inside venv, run ./manage.py test
Pull requests and contributions are welcome. See annotator/static/README.md for more info on frontend architecture.
For help setting up BeaverDam for your application/company, please contact me or leave an issue.
1) Added user-private annotation capability where the annotations made by individual users are visible to them only. Current model had annotations clubbed with the video; but we are moving away from this design and is creating user-groups for every video and annotations shall now be retrieved via a function of (video-id, group-id and user-id) instead of just video-id. 2) Tested and made minor modifications to support image-list where an annotation task can be on a list of JPEG or any image format files instead of video. 3) Dataset shall now be dumped into a file per video (design and implementation in progress)