/YOLOP-TensorRT

unofficial implementation of YOLOP TensorRT

Primary LanguageC++

YOLOP-TensorRT

YOLOP is end2end NN used to detect object and segment drivable and lane segmentations, This is unofficial implementation of YOLOP-TensorRT, whereas the official code was no longer supported by author, and there are still bugs haven't been solved, I reconstruct the code and it is easy to use

most of module like bottleneckCSP etc are from tensorrtx which is great and useful.

update

There are some tiny bugs still exist need to be fixed, still in process.

requirements

  • TensorRT (better >= 7)
  • cuda (better >= 11)
  • a powerful GPU or any Nvidia embedding device like Jetson Nano, NX

installation

  1. firstly modify CMakeLists.txt to adapt your enviroment, like cuda path, openCV
  2. mkdir build & cd build
  3. cmake .. & make

how to use

  1. convert yolop to wts using gen_wts.py, of course you need to download the pytorch model from official yolop
  2. then ./build/buildengine, you can add -w to add your wts file, -b to set batchsize, -o to set output engine
  3. using ./build/inf to inference imgs from directory or videos

output