A13/classification_problems/README.md

Classification Problems Definition

This directory contains prepared data for 2 classification problems with 2 approaches each, derived from the processed skeleton data.

Problem Definitions

Problem A: 3D Classification (Kinect-based)

• Input: 3D joint coordinates from Kinect sensor
• Structure: 13 joints × 3 dimensions (x, y, z) = 39 features per frame
• Temporal: 10 frames per sequence
• Total features: 39 × 10 = 390 features per sequence
• Task: Classify movement quality as Good (1) or Bad (0)

Approaches:

• ADense: Flattened features for dense neural networks
  • Shape: (samples, 390)
• ACNN: Structured features for convolutional neural networks
  • Shape: (samples, 10, 13, 3) - [time_steps, joints, coordinates]

Problem B: 2D Classification (PoseNet-based)

• Input: 2D joint coordinates from PoseNet/MediaPipe
• Structure: 13 joints × 2 dimensions (x, y) = 26 features per frame
• Temporal: 10 frames per sequence
• Total features: 26 × 10 = 260 features per sequence
• Task: Classify movement quality as Good (1) or Bad (0)

Approaches:

• BDense: Flattened features for dense neural networks
  • Shape: (samples, 260)
• BCNN: Structured features for convolutional neural networks
  • Shape: (samples, 10, 13, 2) - [time_steps, joints, coordinates]

Data Organization

The prepared data is organized in the prepared_data/ directory:

ADense Files (3D, Flattened):

• A_Dense_train_X.npy: Training features (shape: samples×390)
• A_Dense_train_y.npy: Training labels
• A_Dense_test_X.npy: Test features (shape: samples×390)
• A_Dense_test_y.npy: Test labels
• A_Dense_train_aug_X.npy: Augmented training features
• A_Dense_train_aug_y.npy: Augmented training labels
• A_Dense_test_aug_X.npy: Augmented test features
• A_Dense_test_aug_y.npy: Augmented test labels

ACNN Files (3D, Structured):

• A_CNN_train_X.npy: Training features (shape: samples×10×13×3)
• A_CNN_train_y.npy: Training labels
• A_CNN_test_X.npy: Test features (shape: samples×10×13×3)
• A_CNN_test_y.npy: Test labels
• A_CNN_train_aug_X.npy: Augmented training features
• A_CNN_train_aug_y.npy: Augmented training labels
• A_CNN_test_aug_X.npy: Augmented test features
• A_CNN_test_aug_y.npy: Augmented test labels

BDense Files (2D, Flattened):

• B_Dense_train_X.npy: Training features (shape: samples×260)
• B_Dense_train_y.npy: Training labels
• B_Dense_test_X.npy: Test features (shape: samples×260)
• B_Dense_test_y.npy: Test labels
• B_Dense_train_aug_X.npy: Augmented training features
• B_Dense_train_aug_y.npy: Augmented training labels
• B_Dense_test_aug_X.npy: Augmented test features
• B_Dense_test_aug_y.npy: Augmented test labels

BCNN Files (2D, Structured):

• B_CNN_train_X.npy: Training features (shape: samples×10×13×2)
• B_CNN_train_y.npy: Training labels
• B_CNN_test_X.npy: Test features (shape: samples×10×13×2)
• B_CNN_test_y.npy: Test labels
• B_CNN_train_aug_X.npy: Augmented training features
• B_CNN_train_aug_y.npy: Augmented training labels
• B_CNN_test_aug_X.npy: Augmented test features
• B_CNN_test_aug_y.npy: Augmented test labels

Data Sources

The data was extracted from the original processed sequences by interpreting the first portion of each frame as joint coordinates:

• For 3D problem: First 39 values per frame interpreted as 13 joints × 3 coordinates
• For 2D problem: First 26 values per frame interpreted as 13 joints × 2 coordinates

Usage Examples

import numpy as np

# Load ADense data for training a dense network
X_train = np.load('prepared_data/A_Dense_train_X.npy')
y_train = np.load('prepared_data/A_Dense_train_y.npy')
X_test = np.load('prepared_data/A_Dense_test_X.npy')
y_test = np.load('prepared_data/A_Dense_test_y.npy')

# Load ACNN data for training a CNN
X_train_cnn = np.load('prepared_data/A_CNN_train_X.npy')
y_train_cnn = np.load('prepared_data/A_CNN_train_y.npy')

Note on Augmented Data

All datasets include both original and augmented versions:

• Original data maintains the original 91 training + 23 test samples
• Augmented data includes 4 additional versions per original sample (mirror, rotate ±10°, stretch)
• Total augmented data: 455 training + 115 test samples