search

linClubs / BEVFusion-ROS-TensorRT

BEVFusion-ROS-TensorRT-CPP real time inference including ros1 & ros2.
BSD 3-Clause "New" or "Revised" License
133 stars 26 forks source link

Regarding lidar2image #2

Open W6WM9M opened 10 months ago

W6WM9M commented 10 months ago

Hello! Thank you for the work.

If I am correct, BEVFusion uses the depth information during inference which is derived from the lidar2image transformation matrix. However, as the lidar and images are taken at different timestamps, the lidar2image transformation matrix shouldn't always be the same since the vehicle is moving.

However, in your code, I do not see a ROS node that subscribes to the vehicle's pose and that the lidar2image is only initialized once in the model. As such, may I kindly check if you assume that the lidar2image transformation matrix is approximately constant?

linClubs commented 10 months ago

hi， The code uses message_filters of ros。the default lidar and camera data have very close timestamps and are considered synchronized. the camera and lidar are rigidly connected and the acquisition time is also synchronized, lidar2cam considered unchanged When the carrier is in motion.