An implementation of the FastAI Object Detection on the Stanford Cars Dataset for the Grab AI Challenge (Computer Vision Category)
Grab AI Challege - Computer Vision Automated Car Detail Recognition PROBLEM STATEMENT: Given a dataset of distinct car images, can we automatically recognize the car model and make?
Larger Objective: How can we help passengers find their rides quicker and in turn increase earning opportunities for our driver-partners by automating the process of digitizing & understanding imagery on the highest possible quality?
Data set: Cars Dataset from Stanford University Details: 18,185 images of 196 classes of cars Training Images: 8,144 Testing Images: 8,041 Classes: Make, Model, Year (e.g., 2012 Tesla model S or 2012 BMW M3 coupe)
The solution is inspired by the Object Detection Technique discussed in the Fast.ai lesson by Jeremy Howard. (Link: https://www.youtube.com/watch?v=Z0ssNAbe81M)
Additional insight was derived from Francesco Pochetti on his implementation on Chest Radiographs (Link: http://francescopochetti.com/detecting-pneumonia-in-chest-radiographs-with-fast-ai/)
All operations are currently limited in the Jupyter Notebook. Future improvements would include:
- Using a larger model (ResNet 50)
- Deployment in a WebApp
Current performance in validation set: 67%