The Polish Sign Language Alphabet Landmark Dataset

The PSL Alphabet Landmark Dataset is a pilot dataset designed for research in Polish Sign Language (PSL) fingerspelling recognition. It provides sequences of hand landmarks extracted from real-time video recordings using the MediaPipe HandLandmarker framework. The dataset supports studies in gesture recognition, sign language recognition, and sequence classification employing machine learning techniques.

To ensure privacy, raw video recordings are not included; instead, anonymized 3D coordinates of hand landmarks are provided. Each frame captures 21 hand landmarks, resulting in 63 features per frame (x, y, z coordinates for each landmark). Frames are grouped into gesture sequences by sample_id, with each sequence consisting of 30 frames, enabling temporal modeling of sign gestures. The dataset encompasses multiple signers, allowing for analysis of inter-signer variability and robustness of recognition models.

Data Format:Each record corresponds to a single frame and includes the following fields:

user_id – unique identifier of the signer

label – gesture or letter label

sample_id – sequence identifier grouping frames into a gesture sample

frame – frame index within the gesture sequence

landmark_{n}_x, landmark_{n}_y, landmark_{n}_z – 3D coordinates of the nth hand landmark

The dataset is structured as a CSV file with one row per frame. Researchers can leverage this structured landmark data for training and evaluating machine learning models for gesture and sign language recognition.

Potential Applications:

Skeleton-based gesture recognition

Sign language recognition and translation

Sequence modeling and temporal pattern analysis

Development of assistive technologies for communication