USL-Suspilne

A small-scale Ukrainian Sign Language dataset for text-to-pose research, built from publicly broadcast news clips on Suspilne Mовлення (Ukrainian public broadcaster). Each clip pairs a Ukrainian sentence with the corresponding interpreter's signing, provided as a video clip and as MediaPipe pose sequences.

Layout

usl-suspilne/
├── README.md
├── train.csv                       # 80% — model training
├── dev.csv                         # 10% — validation / early stopping
├── test.csv                        # 10% — held-out evaluation
├── features/
│   └── v00/{clip_id}.mp4           # H.264, 510×510, 30 fps, signer crop
└── poses/
    ├── mediapipe_holistic/
    │   └── v00/{clip_id}.npy       # (T, 225) — body+hands, 2D + visibility
    └── mediapipe_3d/
        └── v00/{clip_id}.npy       # (T, 150) — 50 joints × (x, y, z)

Split CSVs

Pipe-delimited. One row per clip; header included.

column	meaning
name	{video_id}/{clip_id} — joins to features/, poses/...
text_norm	normalized Ukrainian sentence (input to text→pose models)
signer_id	integer interpreter ID assigned by face clustering (0–4)

text_norm is lowercase Ukrainian with edge punctuation stripped, numerals expanded to words, and % written as відсотків.

Splits

Splits are sampled at the clip level with an 80/10/10 ratio. Every source video and every signer appears in all three splits — this is not a signer- or video-independent benchmark.

Pose formats

Both pose representations are produced by MediaPipe. Frames are aligned with the corresponding .mp4 (30 fps, T = frame count).

mediapipe_holistic

Flattened MediaPipe Holistic output, shape = (T, 225). Reshape to (T, 75, 3) to recover joints:

index range	joints
[0:33]	body (MediaPipe Pose landmarks)
[33:54]	left hand (21 landmarks)
[54:75]	right hand (21 landmarks)

Each joint is (x, y, conf) with x, y normalized to [0, 1] in frame coordinates (origin top-left). The third channel differs by joint type:

• Body joints carry MediaPipe's visibility ∈ [0, 1] (continuous confidence).
• Hand joints carry a placeholder — 1.0 when the hand is detected, 0.0 when it isn't (MediaPipe's hand landmarker doesn't expose a confidence value).

When a body or hand is not detected in a frame, all three values are zero for the corresponding joints. Use that to mask missing landmarks.

mediapipe_3d

3D pose lifted from the 2D Holistic output, shape = (T, 150). Reshape to (T, 50, 3) for (x, y, z):

index range	joints
[0:8]	upper body — Nose, Neck, R/L Shoulder, R/L Elbow, R/L Wrist
[8:29]	left hand (21 landmarks)
[29:50]	right hand (21 landmarks)

Coordinates are clip-relative (not pixel space): x/y are normalized per clip during 2D-to-3D lifting; z is solved under skeletal-length constraints with joint 0 (Nose) anchored at z = 0. Typical value range is roughly [-2.5, 2.5] on each axis.

Source & licensing

Source video belongs to Suspilne Mовлення (UA:PBC) and is redistributed under their terms. Pose features and split CSVs are derivative artefacts of those broadcasts. Verify redistribution rights before publishing this dataset externally.

If you use this dataset, please cite the accompanying thesis (TBD).

Statistics

Auto-generated from the dataset on disk.

split	clips	hours	videos	avg s/clip
train	2731	4.16	17	5.5
dev	342	0.50	17	5.2
test	343	0.52	17	5.4
total	3416	5.17	17	5.5

• Total frames: 558,734
• Signers: 6
• Train vocabulary: 8,120 types / 29,368 tokens
• OOV (token-level): dev 16.6%, test 18.2%