base model training and ego agent

[dempsey-wen]

Hi Yifan,

It is really inspiring to see that you succeeded to align heterogeneous features with a shared pyramid fusion module.

I have two questions:

1. I'm wondering whether the base model has to be a lidar model, or we can also start with a camera image model.
2. Whether the ego can only be lidar or it can also be camera.

I would really appreciate it if you could take the time to answer my questions. :P

[yifanlu0227]

Hi,

1. In my practice, starting with a camera modality as collaboration base is suboptimal to the LiDAR. You could take a look at the open-review forum, where I provided some ablation experiments on it. However, I want to emphasize the worse results may stem from many reasons: the camera detection model is weak / training dataset volume is not enough for camera modality / low image resolution set in the experiments, etc.
2. It can be camera

[dempsey-wen]

Thank you for reply. I've found the answers from the review forum.

Again, it is a very solid work, and thank you for your contribution.