Test task for the RoadAR company.
Code uses opencv for video management and pytorch for model inference.
Before running add the input.avi to the data folder and download the model weights using the config/download_weights.sh script.
All the code can be found in the Task 4.ipynb file.
You can also build a Docker image.
> docker build . --tag test4
> docker run -p 8000:8888 --name test4 test4Then open the localhost:8000 and paste the token from the cmd.
The annotation of the cars can be found in data/input_annotations.json file.
AP(0.5) score of the model equals 46.10%.
Video with the results can be found here.