This repository provides two modular, PyTorch-based pipelines for detecting and aligning human faces in images using MTCNN. One supports roll angle refinement using HopeNet, while the other offers a lightweight version focused solely on eye landmark-based alignment and rotation.
- Detect multiple faces in a single image using
facenet-pytorchMTCNN - Align faces by calculating roll angle from eye landmarks
- Optional: Refine roll angle using HopeNet model
- Handles both full-size images and pre-cropped face images
- Saves aligned faces and logs results in CSV format
- Calculates roll angle based only on eye landmarks
- Automatically rotates and saves aligned faces
- Efficient inference with fallbacks and logging
Install dependencies using:
pip install -r requirements.txtpython face-alig.py --input img-1/ --output results/ --weights hopenet_weights_fixed.pth| Argument | Description |
|---|---|
--input |
Path to input image or directory |
--output |
Directory to save outputs (default: output/) |
--weights |
(Optional) Path to HopeNet weights for roll refinement |
python alignment.py --input path/to/images --output path/to/save --csv results.csv| Argument | Description |
|---|---|
--input |
Directory containing input images |
--output |
Directory to save aligned results |
--csv |
Path to CSV file where alignment results go |
output/aligned_faces/: Aligned face crops (HopeNet version)output/detections/: Face detection visualizationoutput/alignment_results.csv: Metadata with bounding boxes, angles, etc.output/aligned/: Aligned full-face images (lightweight version)results.csv: Angle, inference time, and status per image (lightweight version)
- ResNet50-based model with 3 classification heads for yaw, pitch, and roll.
- Expected to have weights trained on a head pose dataset
- To use HopeNet, download pretrained weights and pass the file path via
--weights
- facenet-pytorch
- HopeNet: Fine-Grained Head Pose Estimation
- OpenCV, PIL, and PyTorch for image processing