pinto0309 / pipnet

The Facial Landmark Preprocessing Toolkit provides efficient preprocessing of facial images and landmarks, supporting image processing, landmark normalization, and data saving for training and testing.

PIPNet Facial Landmark Detection

This repository contains the implementation of PIPNet, a robust approach for facial landmark detection using a deep learning model based on ResNet architectures.

Click here to watch the video

Key Achievements

Exceptional Model Performance on the 300W Dataset

PIPNet model has achieved a significant milestone on the 300W dataset, one of the most challenging benchmarks in facial landmark detection. Successfully attained a minimum Normalized Mean Error (NME) of 2.6%, demonstrating the model's high accuracy and robustness in complex facial recognition tasks.

Features

• Utilizes ResNet as the backbone for the PIPNet model.

• Supports training, testing, and real-time demo modes.

• Includes a 300W dataset loader and loss computation.

• Implements a face detector for real-time landmark detection in videos.

• Designed for easy customization and scalability to accommodate research and development needs.

Requirements

conda create -n PyTorch python=3.8
conda activate PyTorch
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts
pip install opencv-python==4.5.5.64
pip install PyYAML
pip install tqdm

Usage

Datasets: 300W

• Download the datasets from official sources.
• Run below command:

$ python converter.py --data_dir './Dataset'

Training

To train the model, run:

• Configure your dataset path in main.py for training

$ python main.py --train --input-size 256 --batch-size 16 --epochs 60

Testing

For testing the model, use:

• Configure your dataset path in main.py for testing

$ python main.py --test

Real-Time Demo

To run the real-time facial landmark detection:

$ python main.py --demo

Results

Backbone	Epochs	Test NME	Pretrained weights
ResNet18	60	3.39	model

Reference

• https://github.com/jhb86253817/PIPNet
• https://github.com/jahongir7174/PIPNet