- This is our winning submission for the 20-hour Nvidia Hack in AI Hack 2022
- This is a collaboration between me, Mohamed Ali Mimouni and Saif Eddine Layouni
- We took the Building Real-Time Video AI Applications course from Nvidia DLI to learn about TAO and DeppStream then we proceeded to use them for our idea.
- Our presentation
- We used Nvidia's TrafficCamNet and applied transfer learning with TAO to make it detect potholes in images of streets then we supplied it to Deepstream so it applies it in real-time on footage from cars front camera.
- We achieved 57% training accuracy, but it's overfitting and it needs more work for it to be deployable.