FaceMask-RCNN

This repo demonstrates how you can build your own face mask detection mask-rcnn and set up a simple Flask webserver to run inference on real world images.

What we want to achieve:

Step 1: Training the Model

View Step_1_Working_Process.md for steps taken while working for this step.

How to Fine-Tune Mask RCNN with Annotated Face Mask Dataset

Download annotated dataset. Link
Unzip the directory and place the dataset folder (which should contain train and val) .
Create a conda environment: conda create -n face_mask python=3.7.
Activate conda environment: conda activate face_mask
Install requirements: pip install -r requirements.txt.
Training: python train.py --dataset /path/where/you/downloaded/dataset --weights "coco" --training_type "heads" --epochs 5.
Fine tuning: python train.py --dataset /path/where/you/downloaded/dataset --weights "last" --training_type "all" --epochs 10.

Running Evaluation Script

Download trained weights: link
python eval.py --dataset /path/where/you/downloaded/dataset --weights "/path/to/weight/file" --device "cpu"

Training steps, validation and testing

You can take a look at Step_1_Retrain_Mask_RCNN.ipynb for training setps, validation and testing on unseen images.

Step 2: Simple Webserver

Simple Flask webserver which uses Mask RCNN trained to detect and segment face masks in images.

If you don't want to set things up in your computer and just want to try out, here is all-containing co-lab notebook to server a simple webserver.

Preparation

Download weights file: link

How to Run

If you haven't created a conda environment in Step 1, create conda environment: conda create -n face_mask python=3.7
Activate environment: conda activate face_mask
Install required libraries: pip install -r requirements.txt
Start webserver: python webserver.py --weights </path/to/downloaded/weight/file> --device [cpu|gpu]
For local deployment, go to http://127.0.0.1:5000/
- Navigate to /upload and upload an image.
  - You will receive a JSON response which includes the URL of the image.
  - Go to that URL to download the result image.
- Alternatively, you can try out with cURL
```
    curl --location --request POST 'http://localhost:5000/upload' --form 'file=@/path/to/image.jpg'
```

Sample JSON response:

{
  "output":"http://localhost:5000/results/output_image.jpg",
  "percent_masked":1.8,
  "success":true
}

Browser caching: Normally, when you access the image URL, the browser should download the result image. But some browsers open a new tab to show the image. Sometimes, you may see cached result image from previous API call. If that's the case, clear cache and refresh by "Ctrl+Shift+R".

ArkAung/faceMask_rcnn