fDAL-Segmentation

An implementation of Toward Optimal Strategies for Training Self-Driving perception Models in Simulation paper

what has been changed from Narendra's last work?

I made some modifications.

1. As Narendra suggested I reduced the number of classes from 2 to 1 and solved the issues.
2. I changed the model's architecture to be more similar to the original fDAL implementation
3. I changed the loss function to include fdal loss but still have problems with pseudo loss

What are the current issues:

1. no sampling strategies to collect the dataset as the paper suggested
2. no style transfer as the paper suggested
3. pseudo loss makes some problems