PSL-mCFRT: A Multi-View Video Dataset for Continuous Finger-Spelling Recognition and Translation in Pakistan Sign Language

This dataset is a comprehensive, multi-angle collection of continuous finger spelling in Pakistan Sign Language (PSL), developed with the support and guidance of domain experts. Initially, 70 of the most commonly finger-spelled terms used by the Deaf community were identified and documented.Five signers—two males and three females—were recruited to individually finger-spell all selected terms. The recordings were captured using three camera angles: front view, +45° (left), and –45° (right) relative to the center of the signer’s face. During the recording sessions, each signer was filmed from head to knees.To comply with ethical publication guidelines and ensure anonymization, the videos were subsequently processed to conceal the signers’ identities. Specifically, only the region surrounding the signing hand was detected and cropped.The final dataset contains 706 videos. Of these, 353 videos were originally recorded, and the remaining 353 were generated by horizontally flipping (mirroring) the original recordings. All videos contain continuous finger spelling of the identified terms.Dataset FilesPSL_Fingerspelling_Video_Dataset_Cropped.zip (358 MB)Contains the 353 original cropped signing videos.PSL_Fingerspelling_Video_Dataset_Cropped_Mirrored.zip (357.71 MB)Contains the 353 mirrored versions of the original videos.hand_keypoints_features.csv (≈91 MB)Contains frame-level hand landmark coordinates (21 key-points with x–y values) along with extracted geometric features, including inter-finger distances, finger-to-palm distances, thumb interaction measures, and joint-based spatial relationships (e.g., PIP, DIP, and fingertip configurations). These features support machine learning and temporal modelling for continuous fingerspelling recognition.