You are on page 1of 10

Document For AI Training Models

Face Mask Detection


April - 2021

Created By
Rocketflyer Technology Private Limited

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Agenda
This document has been prepared to describe the core practices which have been followed during the training process of AI (Artificial Intelligence) Model -
Face Mask Detection. The document also contains information of architecture used and the overall process flow. The document is capable to provide following
informations-

- Learning Model
- Alert Mechanism
- Data Set
- Steps for training procedure
- Violation detection
- Detection Model
- Core Architecture
- Model Performance Parameter
- Input vs Output

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Learning Model
Deep Learning Model :

We trained a YOLOv3 on our own dataset mAP(Mean Average Precision) on Test Set whereas YOLOv3 achieved 90% mAP on Test
Set even though our test set contained realistic blur images, small + medium + large faces which represent the real world images of
average quality.

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Alert Mechanism
● System monitors the mask, no-mask counts and has 3 status

Safe : When all people are with mask.


Warning : When at least 1 person is without mask.
Danger : When some ratio of people are without mask.

● When model detect any person in frame without mask it will generate alert and save 10 sec alert video, 5 sec before the alert
timing and 5 sec after.

Safe Warning Danger

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Dataset

Set Number of images Objects with mask Objects without mask


Training Set 700 3047 868

Validation Set 100 278 49

Test Set 120 503 156

Total 920 3828 1073

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Process Flow - High Level Technical Flow of Branch
• Load YOLO object detector
Video Ingestion to AI • Filter out weak predictions by ensuring the
detected
Raw Input Video Video Transformation • Calculate count values
Trained AI Model
Source Services • Draw a bounding box rectangle and label
Converted frames on the imagel

AI Processed Output Frame

UI Stream Frames AI Prediction


Processed Output
Frame

Business Inference
Engine

AWS S3
Save Video During Upload To Cloud
Centralized Monitoring Trigger Inferences
inference Occurrence Storage
System

Return URL

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Architecture

● Learning_rate=0.001
● Batch=64
● Subdivisions=32
● Steps=4800,5400

YOLOv3
● Max_batches = 6000
● Approx epochs = (6000*64)/700 = 548
● YOLOv3 Training results: 0.355751 avg loss
● Weights of YOLOv3 trained on Face-mask

Dataset: yolov3_face_mask.weights

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Model Performance

Model Training Set Validation Set Test Set FPS

YOLOv3 99.75% 87.16% 90.18% 23 FPS

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Input vs Output

Input Output

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.
Thank You!
We love solving Engineering and Business problems.If you have a use case in mind which our platform doesn’t cater at present. We would try whatever it takes to
accommodate the suggested feature in our product plan.

Do write us back to us with your feedback, suggestions, queries and custom requirements
Reach out to our Product Engineering Team at product@rocketflyer.in

Strictly Private & Confidential | All rights are reserved by Rocket Flyer Technology Pvt. Ltd.

You might also like