search

InhwanBae / ENet-SAD_Pytorch

Pytorch implementation of "Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)"
MIT License
76 stars 18 forks source link

Question about training #9

Closed jyang68qq closed 3 years ago

jyang68qq commented 3 years ago

Hi, I was digging into your code recently. Nice work!

Just had question about your CULane training dataset. It seems that you didnt use txt which contains axis for lanes in the given dataset.

Could your give a little hint why this would work?

Thanks!

InhwanBae commented 3 years ago

Hi @jyang68qq

Thanks for your interest in my implementation. In the original paper(Section 3.1: Self Attention Distillation), the total loss is comprised of four terms. here, the first _Lseg uses cross-entropy loss, which means the author approaches this lane detection problem as a semantic segmentation task. After predicting 4 lane areas through semantic segmentation, the small network P1 predicts whether each lane really exists or not with 4 binary values.

Following the original paper(Section 3.3: Architecture Design), I use the _laneseglabel provided in the CULane dataset.

I hope this answers your questions.

InhwanBae commented 3 years ago

Thank you for your interest. I'm closing this issue for now. Please feel free to open another issue for other questions.