Kibalama/wlasl-upper-arm-pose-landmarks

WLASL Pose Landmarks (MediaPipe Lite)

This dataset contains 3D pose landmarks extracted from the WLASL (World Level American Sign Language) video dataset using Google MediaPipe. It is designed to facilitate lightweight sign language recognition models by providing pre-computed skeletal data instead of raw video pixels.

Supported Tasks

• Sign Language Recognition (SLR)
• Skeleton-based Action Recognition
• Keypoint Analysis

---

Dataset Structure

The dataset contains the following columns:

Column Name	Type	Description
video_id	string	The original filename of the video (e.g., 001.mp4).
label	string	The English gloss (word) being signed (e.g., "book", "hello").
poses	list	A 3D array-like structure containing the landmark coordinates per frame.

The poses Column

The poses column is a jagged array (variable length) because videos have different durations. Structure: [Frame, Landmark_Index, Coordinate]

• Dimension 1: Frame Index (Variable size $T$)
• Dimension 2: Landmark Index (Fixed size 7)
• Dimension 3: Coordinate (Fixed size 3: [x, y, z])

Keypoint Mapping

The model tracks 7 specific upper-body keypoints relevant to sign language. The indices correspond to the MediaPipe Pose topology:

Index in Array	MediaPipe Index	Body Part
0	0	Nose
1	11	Left Shoulder
2	12	Right Shoulder
3	13	Left Elbow
4	14	Right Elbow
5	15	Left Wrist
6	16	Right Wrist

Note: If a person was not detected in a specific frame, the values for that frame are NaN (Not a Number).

---

Usage

Loading the Dataset

from datasets import load_dataset
import numpy as np

dataset = load_dataset("Kibalama/wlasl-upper-arm-pose-landmarks")

# Inspect the first sample
sample = dataset['train'][0]

print(f"Gloss: {sample['label']}")
print(f"Video ID: {sample['video_id']}")

# Convert poses to numpy array for processing
# Shape: (Frames, 7, 3)
poses = np.array(sample['poses'])
print(f"Pose sequence shape: {poses.shape}")