This project is a Flock-style car tracking program that runs in real time.
I implemented this paper using PyTorch and trained the model on the Veri-776 dataset. You can see my training code in train. Then, I made a simple web server to serve the data over a websocket through an SSH tunnel from the supercomputer to my local device. The system takes advantage of multiple GPUs using the multiprocessing library and shared memory.
I also made a dashboard to visualize it all using the Nuxt.js framework.
implemented model performance:
| Rank-1 | Rank-5 |
|---|---|
| 81.2% | 92.0% |
bugs to fix:
- some streams only load when you refresh the page
- selecting a vehicle in the dashboard turns the simmilarity score on the map on instead of off