The training of the Inverse dynamics model about aco HOT 5 CLOSED

Hi,

1. All the train data in NuScenes is used.
2. I scale the image to 320*180 without center crop.
3. Yes, the interval is 1s.

But note that, the inverse prediction performance is not perfect under this setup (especially throttle prediction, so ACO only uses steering).

from aco.

Thank you for your reply.

from aco.

Hi, currently I get the steer angle from can_bus.vehicle_monitor for nuscenes data. Is this correct? Also, how do you normalize the steer angle to 0-1?

from aco.

I also notice that you give different weight to samples, how to calculate these weights based on the steer?

from aco.

1. Yes, steering is from vehicle monitor.
2. I use statistics (mean, std) from the entire data set for normalization.
3. During training, uniform sampling strategy is used. Different weights are used for visualization and evaluating only.

from aco.