search

XingangPan / SCNN

Spatial CNN for traffic lane detection (AAAI2018)
MIT License
815 stars 242 forks source link

message passing at output layer #79

Closed guozixunnicolas closed 5 years ago

guozixunnicolas commented 5 years ago

Hello, From the paper, there's a comparison between message passing at top hidden layer and at output layer.

May I know where in the output layer did you implement SCNN. Is it at the 3 channel output stage?

Best,

ZX

XingangPan commented 5 years ago

It is at the 5 channel output stage (there are 4 lines and 1 background, so there are 5 channels.)

guozixunnicolas commented 5 years ago

OKay:) How about the feedfoward time in that case? Since I have only one GPU at hand, the GPU runs out memory if I implement message passing at the final stage. So I assume it will take lots of memory and slow down the feed forward time?

XingangPan commented 5 years ago

Applying SCNN at the output requires much less memory than applying to the 128 channel top hidden layer. So you can give it a try. Yes it would slow down the feed forward time, but I can't recall the exact running time.

guozixunnicolas commented 5 years ago

Thank you:)