search

cardwing / Codes-for-IntRA-KD

Inter-Region Affinity Distillation for Road Marking Segmentation (CVPR 2020)
MIT License
114 stars 24 forks source link

Inconsistent F1 of ERFNet on CULane-test (much lower than ERFNet-CULane-PyTorch) #2

Closed Yibin122 closed 4 years ago

Yibin122 commented 4 years ago

Thanks for sharing your latest algorithm!

But I am confused about table 3 in your paper. The F1 measure of ERFNet is only 70.2, which is much lower than 73.1 as previously reported here: https://github.com/cardwing/Codes-for-Lane-Detection

Why is there such an inconsistency?

cardwing commented 4 years ago

The ERFNet model provided in that GitHub repo is trained with several bells and whistles (e.g., warmup pretraining, finetuning, etc). In our CVPR paper, we use the Torch-based ERFNet and it has much lower performance compared with the PyTorch version.

Yibin122 commented 4 years ago

Then did you try guiding the PyTorch version with a better teacher network other than ResNet-101 (72.8)?

cardwing commented 4 years ago

To my knowledge, the PyTorch-based ERFNet model achieves the best F1-measure on CULane-test. I have tried to find a better teacher but failed. You can have a try.