This segmentation network is based on DeepLab network. It is finetuned to detect cars only. A CNN can be connected to it for counting the numbe of cars in the image. The network currently is built for detecting the number of cars.
The dataset can be downloaded from https://fcav.engin.umich.edu/sim-dataset/ . The one used here is the 10K images dataset. THe trained checkpoint is also provided for further finetuning.
The original image -
The segmented image -
Results can be improved by using more data (using the 200k dataset) and running the fine tuning for more epochs.