Kibalama/wlasl-processed-motion-tokens

WLASL Processed Motion Tokens (Normalized)

Dataset Description

This dataset contains normalized, pre-processed 3D skeletal data derived from the WLASL (World Level American Sign Language) dataset.

It is specifically designed for training MotionGPT, VQ-VAEs, and Sign Language Production (SLP) models. Unlike raw MediaPipe output, this data has been geometrically and statistically normalized to remove camera variations, scale differences, and jitter, making it ready for deep learning tokenization.

• Repository: Kibalama/wlasl-processed-motion-tokens
• Source Data: WLASL (MediaPipe Lite)
• Keypoints: 7 Upper-Body Joints (Nose, Shoulders, Elbows, Wrists)
• Feature Space: Position, Velocity, Acceleration (Z-Score Normalized)

Dataset Structure

The dataset consists of video samples where the video pixels have been replaced by high-fidelity motion features.

Columns

Column Name	Type	Description
video_id	string	Original filename (e.g., 005.mp4).
label	string	The English gloss (word) being signed (e.g., "book").
poses	List[List[float]]	Normalized 3D coordinates. Shape: $(T, 7, 3)$.
velocity	List[List[float]]	First-order derivative ($\Delta x$). Shape: $(T, 7, 3)$.
acceleration	List[List[float]]	Second-order derivative ($\Delta^2 x$). Shape: $(T, 7, 3)$.

Data Shape

Each row contains jagged arrays (variable length $T$).

• Dimension 0: Frame Time ($T$)
• Dimension 1: Joint Index (7 joints)
• Dimension 2: Spatial Coordinate (3 values: $x, y, z$)

Joint Indices (MediaPipe Topology)

The 7 keypoints map to the following MediaPipe Pose Landmarker indices:

Index	MediaPipe ID	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

---

Processing Pipeline

The data underwent a rigorous cleaning and normalization pipeline to ensure high-quality motion representation.

1. Data Cleaning

• Interpolation: Missing frames (NaNs) from MediaPipe were filled using linear interpolation.
• Back/Forward Fill: Edge cases at the start/end of videos were filled to prevent gaps.

2. Geometric Normalization

• Per-Frame Centering: In every frame, the midpoint between the shoulders is translated to $(0,0,0)$. This removes the absolute position of the signer in the room.
• Per-Video Scaling: All coordinates are divided by the median shoulder width of the signer across the entire video.
  • Robustness: The median is calculated filtering out the bottom 25% of widths to handle side-views/occlusions.
  • This ensures the signer's size is consistent (scale $\approx 1.0$) regardless of camera distance.

3. Feature Extraction

• Velocity: Calculated using np.diff (pre-padded) to capture motion energy.
• Acceleration: Calculated from velocity to capture force/stops.

4. Statistical Normalization (Z-Score)

The data is standardized using Global Channel Statistics computed from the training set. $$x_{new}=\frac{x-{\mu }_{global}}{{\sigma }_{global}}x\_\{new\}\; =\; \backslash frac\{x\; -\; \backslash mu\_\{global\}\}\{\backslash sigma\_\{global\}\}$$

• $\mu, \sigma$ Calculation: Computed separately for the X, Y, and Z channels (but shared across all joints to preserve relative limb magnitude).
• Result: Values typically range between $[-3, 3]$.

---

Usage

Loading the Dataset

from datasets import load_dataset
import numpy as np

dataset = load_dataset("Kibalama/wlasl-processed-motion-tokens", split="train")

sample = dataset[0]
print(f"Gloss: {sample['label']}")

# Convert to Float32 NumPy Arrays
poses = np.array(sample['poses'], dtype=np.float32) # (Frames, 7, 3)
vels  = np.array(sample['velocity'], dtype=np.float32)
accs  = np.array(sample['acceleration'], dtype=np.float32)

print(f"Pose Shape: {poses.shape}")