Post-editing, a key practice in the translation industry, enables linguists or editors to refine machinetranslated text, enhancing accuracy and efficiency compared to translating from scratch. This approach is also valuable in sign language technology, where it can be applied to oversee and improve the outputs of automatic sign language translation and production systems.
With the rapid growth of AI, there is an increasing need for more accessible and user-friendly interfaces, especially for users without programming skills. This was another key motivation for us to undertake this project.
The interface is divided into two main components: Frame-Level Editing and Segment-Level Editing.
In this section, users can manually edit pose data to correct any inaccuracies from the pose estimator. This functionality is particularly useful for gathering datasets, as it eliminates the need to start from scratch; users can first employ a pose estimator model and then make adjustments. It also aids in evaluating sign pose generation models. Future updates will add features that enable users to adjust the position, orientation, and posture of the signer’s body and hands within the frame.
In the video below, you can see some features of the Frame-Level Editing component of the interface. The user corrects keypoints that were mistakenly placed by the automatic pose estimator.
video1_15fps.mp4
This section allows users to modify the original file—whether it’s a pose or video—at the segment level. If a segment requires adjustment, users can either insert a new video or record one themselves. The model then synthesizes a new video based on the pose sequence of the inserted or recorded video and the signer appearance of the original signer.
segment_level.mp4
A synthetic video sample:
synthetic_video.mp4
Skeleton data:
skeleton.mp4
Installing Packages and Libraries
Use the following command to install all the necessary packages and libraries:
pip install -r requirements.txt
Clone the MusePose repository to your S3IT account using the following command. Since inference with the MusePose model requires a high-end GPU, we recommend running video synthesis on S3IT for optimal performance:
git clone https://github.com/TMElyralab/MusePose.git
Before running the interface, make sure to update the config.yaml
file by changing the path to your desired directory.
To run the interface, use the following command:
python main.py