Notes

1. Why is our SCNN implementation better than the original but appear worse on TuSimple?

If you look at the VGG16-SCNN results on CULane, our re-implementation is no-doubt far superior (over 1.5% improv.), mainly because the original SCNN is in old torch7, while we are based on the modern torchvision ResNets and ImageNet pre-training. We also did a simple grid search for learning rate on the validation set. However, the original paper reports 96.53% on TuSimple, much higher than ours. That is mainly due to it was a competition entry, there is no way we can compete with that.