Human Detection Model

This is a repo that developed a single human detection model(Faster-RCNN) using data from the Motion Keypoint Detection AI Contest conducted by dacon.

Description

See human_detection.ipynb for a description and process of model implementation.

Demo

You can run the model with your images on this site

Repo Structure

.
└── res/
    ├── data                                # Data doesn't go up to the repo.
    └── ...    
└── saved_model/                        
    └── 1                                   # You can download it through the Saved model link below.
└── utils/      
    └── models/                              
        ├── __init__.py 
        ├── base_model.py                   # Pre-trained backbone(resnet50, resnet101, VGG16)
        ├── classifier.py                   # Faster-RCNN's classifier module  
        ├── frcnn.py                        # Faster-RCNN(Integration of all modules)      
        ├── layers.py                       # Other model layer functions(RoI Pooling, NMS, ...)           
        ├── rpn.py                          # Region Proposal Network module
        └── train_step.py                   # Function for model training in 4 steps
    ├── __init__.py                                
    ├── Augmentation.py                     # Data augmentation(Shift, Flip, Add noise, etc...)  
    ├── intersection_over_union.py          # calculate iou between ground truth and prediction                
    ├── label_generator.py                  # A function that generate the ground truth          
    └── utils.py                            # Other necessary functions
├── .dockerignore                           
├── .gitignore                              
├── Dockerfile                              # A file to create an docker image for development environment
├── requirements.txt                        # Libraries needed to create image
├── config.py                               # model config py
├── human_detection.ipynb                   # Examples of progress 
├── SQL_loader.ipynb                        # MySQL communication example
├── README.md                               
└── train.py                                # model training and save weight py

Usage

Data

Download the data through this link

Saved model

link

Development environment

You don't need to clone this repo, just run Dockerfile and the development environment to run this repo image will be installed. And just follow the command line described below. (You can just clone and install requirements and use it.)

docker build --tag {repo image name}:{tag} .

docker network create {network name}
docker run -it [--name {volume container name}] -v {data path}:/data --network {network name} ubuntu /bin/bash
docker run -it -d [--name {mysql name}] -p {sql port}:{sql port} --network {network name} --volumes-from {volume container name} -e [MYSQL_ALLOW_EMPTY_PASSWORD=ture] mysql
docker run -it -d [--name {dev container name}] -p {dev port}:{dev port} --network {network name} {repo image name}:{tag}
docker run -it -d [--name {model server name}] -p {server port1}:{server port1} -p {server port2}:{server port2} --network {network name} -v {saved model path}:/models/{model name} -e MODEL_NAME={model name} tensorflow/serving

Then, you can create a container, connect it, and use it.

If you use MySQL, you can modify train.py and utils/db_uploader.py a little and use it.
Refer to SQL_loader.ipynb

Training

Running train.py will train the model and store the trained weights.
However, since i provide a trained model, there is no need to train it separately.

python train.py

inference

When model training is complete(or received a trained model), modify your {model name} and run inference.py, you can see the result.
Refer to inference.ipynb

python inference.py 

Author

Soon Ho Kwon

github/sooooner
https://tnsgh0101.medium.com/

License

Copyright © 2016 Jon Schlinkert Released under the MIT license.