pedroodb/glosl-isl

GloSL — ISLTranslate / iSign (Indian Sign Language)

Part of GloSL, a set of continuous sign language translation corpora in one format.

• Sign language: Indian Sign Language (ISL) (ins)
• Spoken language: English (en)
• Domain: web videos (news, education, stories) with English text
• Samples: 126,846 (269.03 h) — splits {'train': 101481, 'dev': 12683, 'test': 12682}
• Split policy: glosl-splits-v1, glosl — GloSL-defined (grouped_topic): Source-video-disjoint, topic-stratified: TF-IDF + k-means topic per source video, StratifiedGroupKFold over videos. Replaces the GloSL v0.2 row-level split, where 98% of test clips shared a video with train. (see card.json and diagnostics/)
• License: CC BY-NC 4.0 (README) — public; source: https://github.com/exploration-lab/isltranslate
• Cite: iSign (Joshi et al., Findings of ACL 2024, aclanthology.org/2024.findings-acl.643); ISLTranslate (Findings of ACL 2023, 2023.findings-acl.665); CISLR (EMNLP 2022, 2022.emnlp-main.707); PoseStitch-SLT (EMNLP 2025, 2025.emnlp-main.698); WSLP 2025 overview (2025.wslp-main.1)

Files

annotations.csv, card.json, poses/<NNNN>/<id>.pose. card.json holds the full provenance.

annotations.csv columns (same in every GloSL dataset; empty when not applicable): id, split, text, gloss, gloss_type, signer, start, end, fps, n_frames, video_ref, pose_file, video_file.

Poses

mediapipe-holistic, 543 points (POSE_LANDMARKS 33, FACE_LANDMARKS 468, LEFT_HAND_LANDMARKS 21, RIGHT_HAND_LANDMARKS 21), 3D. Coordinates: pixels of the frame the pose refers to (header width x height, depth 0); z raw. Confidence: body = MediaPipe visibility (1.0 where the source shipped none); face/hands 1 when detected; missing point = (0,0,0) with confidence 0. Frame rate: native, per file (also in annotations.csv). Normalization (glosl-norm-v1) is applied at load time, not stored. Note: reused from slt_datasets MediaPipe-Holistic .npy (KB/12); normalization applied at LOAD time (KB/05)