yifanlu0227 / ChatSim

[CVPR2024 Highlight] Editable Scene Simulation for Autonomous Driving via LLM-Agent Collaboration
https://yifanlu0227.github.io/ChatSim
308 stars 20 forks source link

Why the car can't appear in the camera area? #17

Closed LiuXuan66 closed 5 months ago

LiuXuan66 commented 5 months ago

图片 The prompt is ‘add a black car on the front on the scene‘. I use my own data. I don't know where the problem is?It seems that my camera's posture is incorrect, it cannot face forward on the x-axis.The mcNerf training is right. Is there a fixed code parameter written during the camera generation process?

LiuXuan66 commented 5 months ago

16 Perhaps I misunderstood the role of trans_mat。

yifanlu0227 commented 5 months ago

13 Maybe the same problem? You will have to align your world coordinate's origin with the ground.

LiuXuan66 commented 5 months ago

13 Maybe the same problem? You will have to align your world coordinate's origin with the ground.

Does this mean that the ground is based on the vehicle coordinate system of the first frame? If that's the case, then I have already made the association because my ground feature coordinates are based on the vehicle coordinate system and do not need to be converted like Waymo.

yifanlu0227 commented 5 months ago
Screenshot 2024-04-19 at 21 51 44

Could you check it?

LiuXuan66 commented 5 months ago
Screenshot 2024-04-19 at 21 51 44

Could you check it?

Yes, my own dataset is similar to Waymo, and its car coordinate system is also set at the bottom of the vehicle, with the coordinates facing the same direction as Waymo

LiuXuan66 commented 5 months ago

You can see that the vehicle is generated in the correct world coordinate system, but the coordinates of the camera's blender are obviously incorrect, and its y-axis (triangular tip) is not pointing upwards.

yifanlu0227 commented 5 months ago

I will give you a code snippet to visualize your poses (pose_bounds.npy and cam_meta.npy). I think you should check them first

yifanlu0227 commented 5 months ago

https://github.com/yifanlu0227/ChatSim/blob/main/chatsim/foreground/Blender/utils/blender_utils/utils/visualize_pose.py

yifanlu0227 commented 5 months ago

You need to run it on your PC with a display

LiuXuan66 commented 5 months ago

Thanks a lot. I'll try it.

LiuXuan66 commented 5 months ago

You need to run it on your PC with a display

图片 I have visualized the pose and I am not sure if the world coordinate system in the interface is FLU, so is this pose correct?

yifanlu0227 commented 5 months ago

There is a segment indicating the up direction of your camera, so you can see how your camera is positioned (if you give the correct camera convention)

Screenshot 2024-04-22 at 11 08 03

FLU is a camera convention, not a world coordinate system property. You should know your camera convention in your data format. But at least for now, it's not a normal visualization result. If your dataset is the same as waymo's, with xy plane determining the ground and +z pointing to the ground facing upwards, then probably your camera convention is incorrectly specified, so the cameras are facing crooked, but their trajectory looks reasonable.

v2-77094ec63c1d68a0401cb0f7c10d8faa_r

LiuXuan66 commented 5 months ago

Thank you very much for your answer. I will carefully examine the issue again

LiuXuan66 commented 5 months ago

16 Finally, I obtained the correct pose by annotating trans_mat, and the vehicle can appear within the camera range.