Proper Use of Mask Detection Using Faster R-CNN

For Codig 3.0 Data Mining with theme

The role of data mining to face life in the new normal era

Created by Yakuy 2 Team from Universitas Gadjah Mada

• Ardacandra Subiantoro (18/427572/PA/18532)
• Arief Pujo Arianto (18/430253/PA/18766)
• Chrystian (18/430257/PA/18770)

Abstract

The COVID-19 pandemic has brought many negative impacts for Indonesia and the whole world. The use of masks is an effective way to reduce the spread of COVID-19.We propose the usage of Faster R-CNN model to detect proper wear of masks in photo or video. The model was trained with a dataset of 853 images of groups of people classified into 3 classes: wearing masks, not wearing masks, and using masks incorrectly. The results of the training process resulted in a multi-task loss of 22.4 with a classification speed of 3.10 images per second with GTX 960. The model has been able to detect the correct usage of masks effectively, and we expect it to be used and implemented on CCTV building entry to give a warning if there are people who are not wearing masks or wearing it incorrectly.

Method

Program Implementation

Program Prediction Output

i	Label	Score
1	Without Mask	0.8836
2	With Mask	0.8032
3	With Mask Incorrect	0.2569
4	Without Mask	0.2145

Program Prediction Output with Post-processing.

Inference results with non-maximum suppression.

i	Label	Score
1	Without Mask	0.8836
2	With Mask	0.8032

Webcam demo

Program Results

Proper Mask Detection with Faster R-CNN WebCam Demo
Final Weight (model_with_no_mask.pt) can be downloaded on google drive

Program Performance

Program	Benchmark Time	average inference/s
WebCam Faster R-CNN	61.96	3.15
WebCam Faster R-CNN + Post-processing (nms)	59.30	3.10

Table above shows model performance on GTX 960. We find using simple non-maximum suppression to remove overlapped region did not affect model performance at all.

Poster