Datasets:

Haloroute
/

TriVis-MediaPipe

license: apache-2.0
task_categories:
  - translation
  - video-classification
language:
  - vi
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/shard_*-train.parquet

TriViS-MediaPipe

MediaPipe Holistic 3D keypoints extracted from TriViS, a continuous Vietnamese Sign Language video dataset (~84.5k videos, 12,000 sentences, 3 synchronized camera views).

Dataset overview

Views: every take is recorded by 3 cameras (front, left, right). group4 file names say center for the front camera; the view field always uses the CSV value.
Scenes (metadata field scene):
- lab: 72,453 clips (24,151 takes × 3 views) of 12,000 sentences recorded in a studio, all 4 signer groups. Includes 108 group3 recordings named *_abnormal.mp4, flagged "abnormal": true.
- real: 12,000 clips (4,000 takes × 3 views) of 4,000 sentences recorded in real-life indoor/outdoor settings, 3,000 clips per group.
Annotations: Vietnamese sentence (Sentence), sign-order gloss sequence (Sign_sentence, |-separated) and one of 10 categories.

All clips are in a single train split; build your own train/val/test split from the metadata.

All 4,000 real sentences also appear in lab (same text and id_sentence), and each take has 3 views. To evaluate on unseen sentences, split by id_sentence; to avoid view leakage, keep the 3 views of a take (same scene and id_video) in the same split.

Splits

Split	Labeled rows	Found	Missing	Extracted (uploaded)	Failed / empty
`train`	84,453	—	—	6,000 (extraction in progress)	—

Columns

Column	Type / shape	Description
`pose_pts`	`float32[T, 33, 3]`	Normalized 3D body pose landmarks (`x, y, z`).
`pose_conf`	`float32[T, 33]`	Per-landmark visibility (`0.0` when missing).
`face_pts`	`float32[T, 478, 3]`	Normalized 3D face mesh landmarks (incl. iris).
`face_conf`	`float32[T, 478]`	`1.0` when the face is detected in the frame, else `0.0`.
`left_hand_pts`	`float32[T, 21, 3]`	Normalized 3D left-hand landmarks.
`left_hand_conf`	`float32[T, 21]`	`1.0` when the left hand is detected, else `0.0`.
`right_hand_pts`	`float32[T, 21, 3]`	Normalized 3D right-hand landmarks.
`right_hand_conf`	`float32[T, 21]`	`1.0` when the right hand is detected, else `0.0`.
`transcript`	`string`	Vietnamese sentence (NFC-normalized, whitespace-cleaned).
`fps`	`float32`	Video frame rate (OpenCV).
`duration`	`int32`	Number of decoded frames `T`.
`key`	`string`	`train/<Sentence_video_path>`; unique.
`metadata`	`string`	JSON object, fields below.

Metadata fields

Field	Type	Description
`split`	`string`	Always `"train"`.
`scene`	`string`	`"lab"` or `"real"`.
`filename`	`string`	Video file name.
`video_path`	`string`	`Sentence_video_path` from the label CSV (relative to the TriViS root).
`id_video`	`int`	Take ID (`ID_video`): the same for the 3 views of a take, unique per take within a scene (not across scenes).
`id_sentence`	`int`	Sentence ID (`ID_sentence`); shared between `lab` and `real`.
`group`	`string`	Signer group (`"group1"`–`"group4"`).
`view`	`string`	`"front"`, `"left"` or `"right"`.
`video_belong_id`	`int`	`video_belong_id` column from the label CSV, kept verbatim.
`sentence`	`string`	Cleaned Vietnamese sentence.
`sign_sentence`	`string`	Gloss sequence as in the CSV (`
`glosses`	`list[string]`	`sign_sentence` split on `
`category`	`string`	One of the 10 categories (case variants in the CSVs are unified).
`abnormal`	`bool`	File name ends with `_abnormal.mp4`.
`fps`	`float`	Video frame rate.
`total_frames`	`int`	Number of decoded frames.

Usage

import json
from datasets import load_dataset

ds = load_dataset("Haloroute/TriViS-MediaPipe", split="train")
sample = ds[0]
meta = json.loads(sample["metadata"])
print(sample["transcript"], sample["duration"], meta["scene"], meta["view"], meta["glosses"])