nlitz88 / bevfusion-ros-tensorrt Goto Github PK

Per the output I got here: nlitz88/workzone#15 (comment)

Basically, the model is currently outputting a visualization of the 3D object detection task, but not from the perspective that we want. We want the BEV of the 3D object detections--so we will likely have to spend a bit of time diving into the model's configuration to figure that out.

Basically, get the BEVFusion ROS node running and publishing images. A rough list of subtasks involved in this:

• Configure the environment.sh script in the tool directory to point to the directories containing the headers and library binaries for CUDA Toolkit, libcuDNN, and TensorRT.
  • @CMUBOB97 Got this to work with WSL, and I updated this branch to work if you use the docker image that I'll update this with.
• Update CMakeLists with those similar paths and the lines that they specify
  • Chi also got this to work with his native installation. I decided to just permanently make the additions to the CMakeLists file and use the environment variables set in the environment.sh script.
  • Update the README with more proper instructions as to how to do this--right now it's just a quick note interjected in the middle and is possibly confusing.
• Get the package to build
• Launch the package
• Convert the nuscenes test dataset provided to a rosbag using our fork of their nuscenes to rosbag utility: nlitz88/workzone#15
• Validate output of model when model node ingesting images. I.e., is it outputting a 2D BEV of the 3D object detection? If not, we're going to have to do some (potentially) major surgery to produce the view that we want.
• Test different docker images with different CUDA toolkit and cudnn versions, as well as test with ROS humble (an LTS ROS version)
• Create Docker file with whatever configuration / combination of versions I discover above.