PoseFinder

PoseFinder is a machine learning pipeline for classifying human poses (e.g., "sitting" or "standing") in images using pose estimation and supervised learning. It leverages MediaPipe for extracting pose keypoints and applies a Random Forest classifier to categorize the detected pose.

Features

• Keypoint Extraction: Uses MediaPipe to extract pose keypoints (x, y coordinates) from images.
• Dataset Builder: Constructs a labeled dataset of pose keypoints from a directory of categorized images.
• Model Training: Trains a Random Forest classifier to distinguish between "sitting" and "standing" poses.
• Prediction: Loads an image and predicts the pose using the trained model.
• Export Utilities: Outputs classification results and evaluation reports to Excel.

Repository Structure

.
├── build_dataset.py         # Script to build a dataset from images in 'data/' folder
├── data/                   # Directory for training images, organized by category 
├── export_data.py          # Utility to export classification reports to Excel
├── extract_keypoints.py    # Contains the keypoint extraction logic using MediaPipe
├── model/                  # Directory for trained model artifacts
├── pose_dataset.csv        # Generated dataset of keypoints and labels
├── pose_finder.py          # Main script to predict pose from an input image
├── requirements.txt        # Python dependencies
├── results/                # Directory for results, reports, and evaluation outputs
├── test.jpeg               # Example image for testing
├── train_classifier.py     # Script to train the pose classification model
└── ...

Note: Only a subset of files are listed here. For the full list, see the repository contents.

Installation

1. Clone the repository:

   git clone https://github.com/TaigaTi/PoseFinder.git
   cd PoseFinder

2. Install dependencies:

   pip install -r requirements.txt

Usage

1. Prepare Dataset

Organize your images in the data/ directory as follows:

data/
├── sitting/
│   ├── image1.jpg
│   └── ...
└── standing/
    ├── image2.jpg
    └── ...

2. Build the Dataset

Extract keypoints from images and create a labeled dataset:

python build_dataset.py

This generates pose_dataset.csv.

3. Train the Classifier

Train a Random Forest classifier on the dataset:

python train_classifier.py

• Saves the trained model to model/rf_pose_classifier.joblib
• Outputs evaluation metrics and exports a report to results/results.xlsx

4. Predict Poses

Classify the pose in a test image (default: test.jpeg):

python pose_finder.py

• Prints whether the detected pose is "Sitting" or "Standing".

Key Components

• extract_keypoints.py: Uses MediaPipe to extract pose landmarks for each image.
• build_dataset.py: Loops through labeled images, extracts keypoints, and saves them with labels.
• train_classifier.py: Trains a Random Forest model and saves it for inference.
• pose_finder.py: Loads an image, extracts keypoints, and predicts the pose using the trained model.
• export_data.py: Exports classification reports to Excel for easy analysis.