Open EcustBoy opened 1 year ago
@EcustBoy Thank you for your question. Yes, it is of great help in closed-loop test. Sorry that we do not have the record for that ablation.
@EcustBoy Thank you for your question. Yes, it is of great help in closed-loop test. Sorry that we do not have the record for that ablation.
Thanks for reply~ Dr. chen,so it seems you have done some qualitative evaluation by video record, right? maybe you can provide some qualitative conclusion?
For example, I would like to know if the behavior of stopping at a red light mainly depends on front-view feature refinement design? And if such module is removed, will the model loss the ability to stop at a red light?
I think your conclusion will be helpful for my experiment~ I'll thanks much if you can provide some information
@EcustBoy Yes, we had such observation, thus we drew the conclusion that it was helpful. But we do not stored any qualitative results either.
Hi~ author, I tried to collect carla dataset with my own format by autopilot, and train in open-loop manner, then test my own NN model in carla. But recently I notice it can't learn to stop in front of red light.
In my carla camera sensor setup, I use six 640*320 image with 90 degree FOV for all six views as model input. But I notice in STP3 model, in addition to four-view raw image, a separate front-view image feature is being input to GRU refinement module.
So I wonder whether or not such method is helpful in carla close-loop test?
Is there some ablation results? (it seems your paper only show such abalation experiment in open-loop nuscenes dataset?) And I think if it's really helpful in close-loop test, maybe I can try such method in my model design? (i.e. use separate network to learn front-view image feature)
I'll thanks much for your reply ~~