README.md

title	emoji	colorFrom	colorTo	sdk	app_file	pinned	short_description	license
FacialKeypointDetectionCNN	🙂	red	red	streamlit	src/streamlit_app.py	false	Predict facial keypoints from an uploaded face image	mit

🙂 Facial Keypoints Detection (CNN)

This app predicts facial keypoints (eyes, eyebrows, nose, mouth) from a face image.

🔗 Demo & Code

• 🤗 HuggingFace Space: [https://huggingface.co/spaces/EnYa32/FacialKeypointDetectionCNN]
• 💻 GitHub Repository: [https://github.com/EnYa32/FacialKeypointDetectionCNN]

📊 Visual Results

Predicted Facial Keypoints (Validation Samples)
Model predictions across multiple faces — keypoints align well with facial structures.

---

Training Curves — Loss & RMSE
Stable convergence during CNN training with continuous error reduction.

---

Ground Truth vs Prediction
Predicted keypoints (red) closely match ground truth (green) across most landmarks.

---

RMSE Distribution per Image
Most samples show low spatial error, indicating consistent localization quality.

🎯 Task

Predict landmark coordinates for key facial regions:

• eyes
• eyebrows
• nose
• mouth

Input: face image
Output: 30 regression values (x/y for 15 keypoints)

The app overlays predicted keypoints directly on the processed image.

🧠 Model

• Architecture: CNN (Keras)
• Task type: multi-output regression
• Output size: 30 coordinates
• Loss: regression loss (MSE)
• Input size: 96×96 grayscale

🖥️ App Features

• Image upload (RGB or grayscale)
• Automatic preprocessing
• CNN inference
• Keypoint overlay visualization
• Coordinate table output

Files in this repo

• app.py → Streamlit app
• final_keypoints_cnn.keras → trained Keras model
• target_cols.json → output column names (order of the 30 targets)
• preprocess_config.json → preprocessing settings (image size, normalization)

▶️ Run Locally

pip install -r requirements.txt
streamlit run src/streamlit_app.py

Resize to 96×96

Normalize pixels: x / 255.0

Model predicts normalized coordinates

Convert back to pixel space: y = y * 48 + 48

Clip to valid range: [0, 96]