RecogNet

Yuyang's repo for his PRML (THU Fall 22) course project.

TL;DR

安装依赖

安装必要的依赖：

pip install matplotlib numpy opencv-python pandas seaborn scikit-learn tensorboardX torch torchvision tqdm

并参考 GitHub - DLib 安装 dlib。

准备数据

下载数据集文件，解压到 data/。目录结构如：

data/
├─test_pair/
│  ├─bb.json
│  ├─0/
│  └─...
└─training_set/
│  ├─bb.json
   ├─Alice/
   ├─Bob/
   └─...

训练模型

python 1_train.py

测试模型

我们提供一个训练好的 ResNet34 的 checkpoint。

python 1_test.py --checkpoint [PATH_TO_CHECKPOINT]

测试数据预测结果

见 test_res.txt

Environment Setup

Install dependencies from pip

pip install -r requirements.txt

`dlib` (required if you run data preprocessing)

We utilize dlib for face detection in data preprocessing.

Build from Source

First, cd into thirdparty/, and clone the repository:

git clone https://github.com/davisking/dlib.git

Then, cd into dlib/, and build from source:

mkdir build
cd build
cmake .. -DUSE_AVX_INSTRUCTIONS=1
cmake --build .
cd ..
python setup.py install --set USE_AVX_INSTRUCTIONS=1

If you have a CUDA compatible GPU, you can build with GPU support:
mkdir build
cd build
cmake .. -DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1
cmake --build .
cd ..
python setup.py install --set USE_AVX_INSTRUCTIONS=1 --set DLIB_USE_CUDA=1
This requires CUDA and cuDNN library.

Download Face Landmark Predictor

Download predictor into ckpt/

Data Preparation (Optional)

Download the dataset zip file, unzip it in data/. The directory should be like:

data/
├─test_pair/
│  ├─bb.json
│  ├─0/
│  └─...
└─training_set/
│  ├─bb.json
   ├─Alice/
   ├─Bob/
   └─...

Image Pre-processing

To help training, we first pre-process the image:

rotate the image, so that the eyes are on a horizontal line.
detect the face and mark the bounding box

Pre-process the image:

python data-proc.py --dataset_dir data/

This will generate 2 bounding boxes for each image: one for the untransformed version, and another one for the transformed version. Parameters for the affine tranformation as well as the bounding boxes are in data/training_set/bb.json or data/test_pair/bb.json.

On our provided images, at least one face is (will be) detectable with our code. If you want to use your images, you may need to check that there is no failure cases, or you should optimize the detection code or manually annotate the data.

Train the model

Specify the training parameters in config/hp.yaml, then run:

python 1_train.py

We also provide a checkpoint recognet_demo_ckpt.pt (PENDING). Download it and put it to ckpt/recognet_demo_ckpt.pt.

Test the model

Specify the hyperparameters in config/hp.yaml, then run:

python 1_test.py --checkpoint [PATH_TO_CHECKPOINT]

YuyangLee/FaceRecog