The Carla simulation tool is a nice way to generate groundtruth to automatically get bounding box positions and dimensions.
Once you have Carla installed, you can launch it with:
./CarlaUE4.sh -RenderOffScreen
Then you the notebook script to spawn vehicles.
You can spawn vehicles multiple times to constantly have cars driving by.
Carla simulation
This will save your images and the associated labels (X-Y positions and width/height/length of each vehicles). This is how you make your dataset.
I recommend to have the same camera pose and focal length to have the most realistic simulation.
Due to a bug with Carla 0.9.14, some vehicles like motorbikes dont have correct dimensions. This is why my neural network weights will not work with two-wheelers...
You can visualize your annotation by using visualize_carla_annotation.py
Visualize bounding boxes
