GibranBenitez / FaceMask_Twitter

Twitter Face Image Mining for Recognition of Different Face Mask Types

1. Intro
2. Requirements
3. General Instructions
4. Preliminary Dataset
5. Face Mask Recognition

Introduction

• This is the repository of our paper Twitter Face Image Mining for Recognition of Different Face Mask Types
• Follow the general instructions to run the complete process for Face Mask Image Mining on Twitter
• If you are interested in the preliminary dataset download it here (460MB)
• If you want to try the face mask recognition model only, follow these instructions

Requirements

• Install the necessary libraries:
  • wget
  • tweepy
  • PyTorch version 1.10.0+ and torchvision 0.3.0+

Face Mask Image Mining

• Clone this repository $ git clone https://github.com/GibranBenitez/FaceMask_Twitter.git
• Update your Twitter API V2 Academic Research access level credentials in ./MaskTwitter/credentials.py
• Download the ConvNeXt weights (190MB) on ./MaskClassify/weights/
• Modify the date range and keywords of faceMaskTwitter.py and demo_covid_classy.py and run these commands.

$ python ./RetinaFace/faceMaskTwitter.py
$ python ./MaskClassify/demo_covid_classy.py

• Example values for faceMaskTwitter.py (keywords and date range):

keywords = ["n95","ffp2","face mask","cubrebocas","barbijo"]
dates = ["2019-12-01","2019-12-31"]

• Example values for demo_covid_classy.py (keywords and folder name of dates):

keywords = ["n95","ffp2","face mask","cubrebocas","barbijo"]
 dates = "2019-12-01__2022-12-31"

Preliminary Dataset

• We manually annotate 10,500 images based on the four face mask types: surgical masks, cloth masks, respirators, and valved mask, as well as unmasked faces.
• You can download the dataset from Google Drive (460MB)
• The number of images for train, validation, and test sets are as follows

Set	Surgical	Cloth	Respirator	Valved	Unmasked	Total
Train	2200	1700	1000	1000	2100	8000
Val.	200	200	200	100	300	1000
Test	350	330	300	170	350	1500

Face Mask Recognition

• We train a ConvNeXt architecture with our preliminary dataset. The results obtained with validation and test sets are 94.8% and 94.1%, respectively.
• To run a classification demo you have to download the ConvNeXt weights (190MB) on ./MaskClassify/weights/
• Modify the data path data_dir, the output folder out_dir, and run $ python ./MaskClassify/demo_covid_classy.py
• Example of modifications from demo_covid_classy.py (data_dir, includes 10 demo images):

############## Uncomment this section for demo classy only ################    
dates = None
############## Chose pahts for demo classy only ###########################
data_dir = "./MaskClassify/demo_imgs"   
out_dir = "./MaskClassify/output"