BIM-SIGN Pose: Malaysian Sign Language Skeletal Pose Dataset

BIM-SIGN Pose Data : Malaysian Sign Language Isolated-Sign Skeletal Dataset BIM-SIGN (Pose) provides MediaPipe Holistic skeletal pose sequences for the largest isolated word-level Malaysian Sign Language (Bahasa Isyarat Malaysia, BIM) video corpus assembled to date. It is the pose-only companion of the BIM-SIGN dataset, containing no video or imagery, only numerical skeletal coordinates extracted from the source recordings, making it suitable for open, non-identifying distribution. The dataset covers 117 gloss classes spanning everyday communication (greetings, family terms, food, transport, question words, common verbs) and includes pose sequences for 15,266 of 15,277 recorded clips (99.9% coverage), contributed by approximately 220 signers, with 92 to 404 samples per gloss (mean ~130). Each sample is provided in three complementary representations, extracted with MediaPipe Holistic (min_detection_confidence=0.5, min_tracking_confidence=0.5): landmarks_full_1662 — raw Holistic output (T, 1662): full body pose (33 points, x/y/z/visibility), face (468 points), and both hands (21 points each), in MediaPipe-normalized [0,1] coordinates. normalized_1662 — the same layout after anchor-aware normalization: the skeleton is centered at the neck and scaled by neck-to-nose distance for camera-distance invariance, with each hand re-centered on its own wrist so handshape and hand location are represented as separable factors. keypoints_258 — a compact body-and-hands-only representation (T, 258), excluding the face, for lighter-weight modeling. Sequence lengths range from 5 to 712 frames (mean 70.6, median 66). A seeded, stratified 80/10/10 train/validation/test split (seed 42) is provided, along with a full gloss vocabulary with English translations, per-gloss statistics, and a manifest documenting per-sample data availability and known quality-control flags (a small number of empty-pose or length-outlier clips are disclosed and itemized). The dataset is intended to support isolated sign language recognition, sign language production (gloss-to-pose synthesis), signer-style modeling, and pose-based retrieval research for a low-resource sign language. A companion package containing the raw RGB video recordings from which these poses were extracted is published separately, given its identifiable content; the two packages share sample identifiers and can be merged by researchers with access to both.