ViSignLanguage-MediaPipe

This dataset holds MediaPipe Holistic 3D keypoints extracted from the ViSignLanguage dataset of Vietnamese isolated-sign videos. Each row is one video clip. The rows contain per-frame landmarks for the body pose, the face mesh and both hands, so models can be trained on keypoints without decoding any video.

• Task: isolated sign recognition with 100 classes.
• Clips: 7,295 short clips, each 16–30 frames long, from 224×224 source videos.
• Format: Parquet (ZSTD), with tensor columns stored as Hugging Face Array2D/Array3D features.

Splits

Split	Clips	Shards	Labels	Parquet size
train	2,806	8	✅ 100 classes (5–52 clips per class)	~480 MB
public_test	1,630	4	❌ unlabeled	~290 MB
private_test	2,859	6	❌ unlabeled	~504 MB
Total	7,295	18		~1.27 GB

• Shards are named data/shard_<N>-<split>.parquet. The YAML header maps them to the split names, so load_dataset returns all three splits with no extra arguments.
• Every source video was extracted successfully: no split has failed or empty clips.
• Deduplication (2026-11-14): the original train split had 3,875 clips. 1,069 of them were duplicates, named <id>_<n>.mp4 (for example 13326_1.mp4), whose base clip <id>.mp4 was also present. Their keypoints were identical to the base clip's, so they were removed. No filename or key in the dataset ends in _<digits> any more. The test splits contained no duplicates.
• The source dataset publishes no labels for the test splits. In those splits transcript is "" and label_id is null.
• Train shards hold 268–370 rows each, and test shards hold up to 500.

Schema

T is the clip's frame count (its duration, 16–30).

Column	Feature type	Shape	Description
pose_pts	Array3D(float32)	(T, 33, 3)	Body pose landmarks (x, y, z)
pose_conf	Array2D(float32)	(T, 33)	Per-landmark visibility from MediaPipe Pose, in [0, 1]
face_pts	Array3D(float32)	(T, 478, 3)	Face mesh landmarks (x, y, z), including iris points
face_conf	Array2D(float32)	(T, 478)	1.0 if a face was detected in the frame, else 0.0
left_hand_pts	Array3D(float32)	(T, 21, 3)	Left-hand landmarks (x, y, z)
left_hand_conf	Array2D(float32)	(T, 21)	1.0 if the left hand was detected in the frame, else 0.0
right_hand_pts	Array3D(float32)	(T, 21, 3)	Right-hand landmarks (x, y, z)
right_hand_conf	Array2D(float32)	(T, 21)	1.0 if the right hand was detected in the frame, else 0.0
transcript	string	–	Gloss (class label), NFC-normalized Vietnamese. "" in the test splits
fps	float32	–	Video frame rate, probed with OpenCV (the metadata CSV is the fallback)
duration	int32	–	Number of decoded frames, T
key	string	–	Unique ID: train/<label>/<filename> or <split>/<filename>
metadata	string	–	JSON object; see below

metadata JSON

Field	train	Test splits	Description
label	✅	""	Gloss, NFC-normalized (same as transcript)
label_id	✅	null	Integer class ID from label_mapping.json
split	✅	✅	train, public_test or private_test
filename	✅	✅	Source video filename, e.g. 133326.mp4
size_mb, fps, total_frames, width, height, duration_s	✅	–	Copied from the source dataset_metadata.csv (train only)

Example from train:

{"label": "An ủi", "label_id": 0, "split": "train", "filename": "133326.mp4", "size_mb": 0.021, "fps": 20.01, "total_frames": 30, "width": 224, "height": 224, "duration_s": 1.499}

Keypoint layout and missing data

• Pose: 33 landmarks. x and y are normalized image coordinates in [0, 1], and z is relative depth with the hips as origin. Confidence is MediaPipe's per-landmark visibility.
• Face: 478 landmarks (the 468-point mesh plus 10 iris points), with normalized (x, y, z). Confidence is binary.
• Hands: 21 landmarks per hand, with normalized (x, y, z). Confidence is binary.
• Missing detections: when a part is not detected in a frame, all of its points for that frame are 0.0 and its confidence is 0.0. No NaN values are used. Mask with *_conf > 0 rather than checking the coordinates. Hands are often out of frame: in a sample, the left hand was missing in about 40–70% of frames and the right hand in about 30–40%.
• The extractor's leading person axis (1, T, …) was squeezed out, giving (T, …).

Labels

There are 100 classes. The full label -> id mapping (NFC-normalized) is in label_mapping.json. The table below gives the number of train clips per class after deduplication.

All 100 classes

ID	Label	Train clips
0	An ủi	24
1	Áp dụng	50
2	Ăn	43
3	Ăn mừng	47
4	Ban ngày	15
5	Ban đêm	5
6	Bàn tay	5
7	Băn khoăn	43
8	Bạn thân	15
9	Bế mạc	26
10	Bệnh nhân	5
11	Bệnh viện	36
12	Biết	46
13	Biếu tặng	38
14	Bộ y tế	14
15	Cá	41
16	Cách ly	5
17	Cám dỗ	49
18	Cần	6
19	Cảm ơn	46
20	Chào	26
21	Chân	48
22	Chấp nhận	5
23	Chạy	51
24	Chậm lại	49
25	Chiều	45
26	Cho	34
27	Chúng ta	12
28	Con gấu	41
29	Có thể	15
30	Cơ thể	5
31	Cứu	37
32	Dạy dỗ	30
33	Dễ	35
34	Ghét	23
35	Giúp	30
36	Hâm mộ	41
37	Ho	5
38	Hôm nay	17
39	Họ	35
40	Học sinh	5
41	Hy sinh	37
42	Kết hôn	23
43	Khai báo	40
44	Khẩu trang	8
45	Khóc	45
46	Khu cách ly	9
47	Lây bệnh	36
48	Lo lắng	5
49	Mời vào	40
50	Mua	26
51	Nặng	34
52	Nghe	49
53	Nghỉ ngơi	49
54	Ngón tay	5
55	Nhà	51
56	Nhầm	5
57	Nhìn	40
58	Nhớ	40
59	Nói	21
60	Nói xấu	50
61	Nôn ói	31
62	Ô tô	33
63	Phạt	26
64	Phía sau	49
65	Phỏng vấn	32
66	Phục hồi	5
67	Rau	51
68	Rẽ phải	12
69	Rẽ trái	47
70	San sẻ	33
71	Sốt	5
72	Sử dụng	44
73	Tập luyện	9
74	Thất lạc	20
75	Thăm	13
76	Thích	8
77	Thương	5
78	Thức ăn	39
79	Thức dậy	51
80	Tôi	36
81	Tối	18
82	Trưa	41
83	Trường học	52
84	Uống	19
85	Ủng hộ	5
86	Vâng lời	5
87	Virus	5
88	Xa	42
89	Xe máy	5
90	Xe đạp	13
91	Xin lỗi	14
92	Xin phép	28
93	Xuất viện	35
94	Xúc động	43
95	Đâu	39
96	Đầu	36
97	Đẹp	14
98	Đi	46
99	Đồng ý	36

Usage

import json
import numpy as np
from datasets import load_dataset

ds = load_dataset("Haloroute/ViSignLanguage-MediaPipe")
print(ds)  # train: 2806, public_test: 1630, private_test: 2859

sample = ds["train"].with_format("numpy")[0]
print(sample["key"], sample["transcript"], sample["duration"])
print(sample["right_hand_pts"].shape)        # (T, 21, 3)

meta = json.loads(sample["metadata"])
label_id = meta["label_id"]

# Frames where the right hand was detected
rh_valid = sample["right_hand_conf"][:, 0] > 0

Extraction

The keypoints were extracted with the MediaPipe Tasks Holistic Landmarker (holistic_landmarker.task) and written as one Parquet row per video. Labels come from the source folder names and were mapped to IDs with the source label_mapping.pkl. Train metadata was joined from dataset_metadata.csv on filename, because that CSV's label and filepath columns are mojibake.