Nonovogo/face-emotions2

MediaPipe Face Emotion Classifier

This model classifies facial expressions into 7 categories using MediaPipe Face Landmarker Blendshapes and a Random Forest Classifier.

Performance Summary

This model was trained on the FER-2013 dataset. Given the difficulty and noise of $48 \times 48$ grayscale images, the model achieves a strong baseline by utilizing high-level muscle movement features (Blendshapes).

• Overall Accuracy: 51.36%
• Baseline (Random Guess): 14.28%
• Human Baseline (FER-2013): ~65-68%

Detailed Metrics

Emotion	Precision	Recall	F1-Score
Happy	0.66	0.80	0.72
Surprise	0.68	0.71	0.70
Disgust	0.71	0.50	0.59
Neutral	0.41	0.48	0.44
Angry	0.40	0.41	0.41
Fear	0.38	0.27	0.32
Sad	0.32	0.30	0.31

[Image of a Confusion Matrix for facial emotion recognition]

Model Architecture

1. Feature Extraction

Instead of raw pixel data, we use the 52 Blendshape scores provided by MediaPipe. These scores represent specific facial muscle activations (e.g., browInnerUp, mouthSmileLeft). This makes the model robust to different lighting conditions and head poses.

2. Classifier

• Algorithm: Random Forest Classifier
• Estimators: 200
• Optimization: Trained with class_weight='balanced' to ensure fair representation of rare emotions like Disgust.

How to use

You can load this model using the skops library:

import skops.io as sio

# Load the model
model = sio.load("model.skops", trusted=True)

# Expects an array of 52 blendshape scores
# prediction = model.predict([blendshape_list])