retinaface4j

Because not all production environements can be in python

RetinaFace (RetinaFace: Single-stage Dense Face Localisation in the Wild, published in 2019), in java.
This repo is an experiment attempting to answer the question : Is pytorch java + ND4J a viable option for deep learning on the JVM ?

Original paper -> arXiv
Original Mxnet implementation -> Insightface
Insprired by the Pytorch implementation -> Pytorch_Retinaface

Table of contents

1. Installation
2. Usage
3. Acknowledgements

example output :

---

INSTALLATION

• Download and unpack libtorch from the pytorch home page (or 1.4 or greater). From the pytorch.org homepage under "Quick Start Locally", make sure "LibTorch" is the selected package.
• Run export LIBTORCH_HOME=/path/to/libtorch. The build.gradle file will use this to set java.library.path when running the application. If you are using PyTorch in your own environment, LIBTORCH_HOME is not necessary. Instead, you will need to set java.library.path to /path/to/libtorch/lib.

USAGE

Download pretrained weights on Dropbox , and save them in the src/main/resources/models/ folder

Run :

./gradlew run --args="./sample-images/WC_FR.jpeg output.png"

Java usage :

import java.util.*;
import javax.imageio.ImageIO;
import java.io.File;
import java.io.IOException;
import retinaface4j.Detector;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.datavec.image.loader.NativeImageLoader;

String modelPath = "./models/Retinaface_resnet_traced.pt";
String imgPath = "./sample-images/WC_FR.jpeg";
Double detThresh = 0.9;
Double nmsThresh = 0.4;
BufferedImage img = null;
INDArray nd4jimg = null;
try {
    img = ImageIO.read(new File(imgPath));
    NativeImageLoader loader = new NativeImageLoader(img.getHeight(), img.getWidth(), 3);
    nd4jimg = loader.asMatrix(img);
} catch (IOException e) {
    System.out.println("[ERROR] could not read image");
    System.out.println(e.getMessage());
}

Detector detector = new Detector(modelPath, detThresh, nmsThresh);
INDArray dets = detector.predict(nd4jimg);

ACKNOWLEDGEMENTS

This work is laergely based on :
The original implementation by the insightface team.
The pyton pytorch implementation.
Pytorch's java-demo example lib.
If you use this repo, please reference the original work :

@inproceedings{Deng2020CVPR,
title = {RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild},
author = {Deng, Jiankang and Guo, Jia and Ververas, Evangelos and Kotsia, Irene and Zafeiriou, Stefanos},
booktitle = {CVPR},
year = {2020}
}