nv-tlabs/lift-splat-shoot

Training with drivable areas / maps?

yzhou377 opened this issue · 7 comments

yzhou377 commented

Hi!

Thanks for your great contribution! I was trying to training the network again with NuScenes and other datasets, but I noticed that the train.py code does not incorporate the drivable area information, which supposedly should come from the map. Will this functionality be provided? Thanks!

Best,
Yiyang

pha-nguyen commented

@yzhou377 thanks for raising the same concern, I am also waiting for two missing modules which are mentioned in the paper. Those are drivable area and lane segmentation. Have you tried to implement it yourself, please share any thoughts or progresses if it doesn't bother you, appreciate pretty much!!!

yzhou377 commented

I have implemented the drivable area part myself on a different dataset. Not with NuScenes. You may consider adding one more channel at the output for the drivable area part.

zhangchbin commented

Hi, @yzhou377 and @pha-nguyen ,
I also implemented the drivable area and road on the Nuscenes, but the performance is worse.
Do you have the same trouble?

  • 👍1
ningerlove commented

who can provide drivable area and lane segmentation code

gao-shi commented

I added drivable area segmentation task to the network， and 100m * 50m setting IOU is about 79, 30m * 30m is about 89, absolutely it works, but the code cannot be shared due to company policy

  • 👍2
JiweiMaster commented

I added drivable area segmentation task to the network， and 100m * 50m setting IOU is about 79, 30m * 30m is about 89, absolutely it works, but the code cannot be shared due to company policy

hello，Can you provide some ideas for generating maps？ I'm new to nuScense.

linlion0311 commented

我在網絡中添加了可行駛區域分割任務，100m * 50m設置IOU約為79，30m * 30m約為89，絕對可以，但由於公司政策不能共享代碼

Hi, can you tell us some tricks for generating maps task？ thanks~