Traffic Sign Recognition

Codebase For Team CSPDarknet53's Interpretation Of The AI In Geoservices Challenge From The Artificial Intelligence International Junior Contest

Contents

• Code

• Website Demo

• Details

Code

Model Pipeline & Training:

• EfficientNetB2.ipynb : Notebook To Bring Together The Pipeline

• Pipeline :

  • constants.py : Some Constant Values
  • dataset.py : Data Preprocessing / Loading
  • generate_submission.py : Generates Submissions For The Contest
  • models.py : Model Dictionary & Class
  • train.py : Model Training Module
  • utils.py : Helper Functions

Website :

• inference.py : Helper Function For Flask App

• main.py : Flask App Backend

• App :

  • Test Images (From WikiPedia)
  • Class2Label Dictionary
  • English Conversion For Classes
  • Russian Conversion For Classes

• Frontend Code :

  • index.html : FrontEnd For Website HomePage
  • result.html : FrontEnd For Predictions

• Styling :

  • Images For Website
  • Styles For Website

Website Demo

AIIJC.Website.Demo.mp4

Details

Model

Pre-Trained Model : EfficientNetB0 With Custom Head

Image Size: (224, 224)

Image Preprocessing: pseudo-labelling, albumentations

Training Time: 14 hours

Training Accuracy: 99.86%

Validation Accuracy: 99.72%

Testing Accuracy: 99.2%