Sign Language Tools

Sign Language Tools is a Python library that provides the building blocks needed to work with sign language datasets made of videos, pose landmarks (from MediaPipe or OpenPose) and their temporal annotations (glosses, lemmas, sign boundaries, etc.).

It was originally developed by the LSFB team at the University of Namur to build the machine-learning pipeline behind the LSFB dataset, and was later generalized to work with other sign language datasets.

The library is designed to pair with sign-language-data-loading, which loads poses, videos and annotations from a dataset on disk. sign-language-data-loading gets you a sample into memory; sign-language-tools is what you use to extract, transform, visualize and play it back. See Working with sign-language-data-loading for a full example combining the two.

What's included

• Pose — Extract MediaPipe holistic landmarks from video, connectivity/edge definitions for MediaPipe and OpenPose skeletons, data-augmentation and preprocessing transforms (interpolation, resampling, smoothing, normalization, cropping, ...), and matplotlib plotting helpers for landmarks.
• Video — Decode video files frame by frame, and PyTorch-compatible transforms for video tensors (temporal crop, temporal padding).
• Video Player — An OpenCV-based, modular player to visualize a video alongside its pose landmarks, annotation segments, time series or heatmaps, all synchronized on a single playback clock.
• Annotations — Transforms to manipulate segment-based annotations (merging, scaling, converting to frame labels or BIO tags, ...), plus IoU/NMS utilities and a timeline plotting helper.
• Common Transforms — Small, data-agnostic transform building blocks (Compose, Concatenate, Randomize, ...) used to build transform pipelines across the other submodules.

Project status

This project is under active development. Contributions and issues are welcome on GitHub.