autonomousvision / carla_garage

[ICCV'23] Hidden Biases of End-to-End Driving Models
MIT License
203 stars 16 forks source link

The question about plant #34

Closed raozhongyu closed 4 months ago

raozhongyu commented 5 months ago

I notice that the input route and the output checkpoint ground truth is same. How it works?

Kait0 commented 5 months ago

The checkpoints predict the route centerline, they are the same. Since the route centerline is available in the HD-map there is no reason to predict it, one could also directly use it (e.g.: like in map_agent). We predict it anyway with the disentangled representation in this plant implementation. There is no particular reason for it. In some sense it makes sure that the architecture has no information bottleneck that prevents it from copying its input.

raozhongyu commented 5 months ago

Thanks a lot. I get it. Have you tried only predict the target speed prediction?

Kait0 commented 5 months ago

no