AI Hand Gesture Recognition (concise)

WHITEPAPER: https://drive.google.com/file/d/1C1uuupBMZ_RwiGJ-QhDkHX9WDe_hndD7/view?usp=sharing

Lightweight pipeline to extract MediaPipe hand landmarks from hand images and train classical ML models (logistic regression, SVM, and MLP) to predict raised finger count (0–5).

What’s in this repo

• annotate_landmarks.py — extract 21 MediaPipe landmarks (x,y,z) per image and save CSV rows.
• verify_landmarks.py — visual check: overlay landmarks on images (random samples).
• logistic_regression_fingers.py — logistic regression baseline: train/val/test split, hyperparameter sweep, confusion matrix, CV, learning curve, permuted-label sanity check.
• svm_fingers.py — SVM utilities and evaluation helpers (grid search, plots, CI calculation).
• mlp_finger_classifier.py — MLP baseline with similar evaluation flow to logistic regression.
• landmarks*.csv — example CSV files produced by the annotation pipeline (feature columns: x0..x20, y0..y20, z0..z20, label column: label_fingers).

Quick start

1. Install deps:

pip install -r requirements.txt

2. Extract landmarks from an images folder:

python annotate_landmarks.py --input path/to/images --output landmarks_reduced.csv

3. Verify a few annotations:

python verify_landmarks.py

4. Run the logistic regression baseline (example):

python logistic_regression_fingers.py

Notes: the model scripts expect CSVs with image_path, label_fingers, x0..x20, y0..y20, z0..z20. Adjust CSV_PATH in the scripts if needed.

Outputs and evaluation

• Confusion matrices, learning-curve CSVs, and CV summaries are produced by the model scripts.
• Files written by the scripts (e.g., logistic_regression_learning_curve.csv, mlp_learning_curve.csv) are saved in the repository root by default.

Tips

• Use label_fingers values 0–5. Many scripts filter to supported labels via their LABELS config.
• If you change dataset filenames or column layout, update the feature_cols selection in the model scripts.

Contact / Notes

This repo is focused on landmark-based classical ML baselines (not end-to-end image CNNs). If you want a script added (e.g., a single train_models.py orchestrator), I can add one.