search

autonomousvision / carla_garage

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

Assert Image_seq_len==1 #6

Closed raozhongyu closed 1 year ago

raozhongyu commented 1 year ago

I notice that the Image_seq_len is assert as 1. I'd like to konw if the author have try the image_seq_len as other number and what's the performance. Meanwhile, I also like to konw if I want change the image_seq_len, what should I notice?

Kait0 commented 1 year ago

We have not tried using multiple image frames and it is not implemented (that is why the assert exists). We have tested multiple LiDARs as described here. If you want to change in image_seq_len you need to change data.py, your backbone architectures, and sensor_agent.py accordingly. You can look at how lidar_seq_len works for an example.

raozhongyu commented 1 year ago

Thanks a lot.

raozhongyu commented 11 months ago

Hi, I notice that the bev semantic prediction is in range [64, 64] which is same as the lidar resolution. But the paper only use the front image, and the paper says that the semantic prediction is only based on the visible area. I don't find how to convert the bev semantic label to the visible area? Could you give me some suggestions.

Kait0 commented 11 months ago

A mask is applied to the regions not visible in the camera.