This repository is a full python ros package that is made similar to the darknet_ros with only difference that it uses tensorrt acceleration.
So, it is faster on devices like jetson nano, rtx2080, tx2 or xavier.
And it is easier to maintain because fully written in python.
First you need to install tensorrt and opencv-python. Those are platform specific and a bit tricky to isntall, so I didn't put those in requirements.txt.
- To get opencv run
python -m pip opencv-pythonor build it yourself. - Installation of tensorrt is also easy: download appropriate tensorrt version from official webiste tensorrt_download_link ("download now" on webiste is a lie - you need to login) and run:
python -m pip install $TENSORRT_PATH/python/my_trt_version_cp27_none_linux_x86_x64.whl
- If you are using global python shame on you, use something like pyenv + pyenv-virtualenv instead, but luckily you don't need to do anything. If you are using virtual environment - link tensorrt to your site-packages like this:
$ ln -s /usr/local/lib/python2.7/dist-packages/tensorrt $VIRTUAL_ENV/lib/site-packages/- For opencv I suggest build it yourself, check buildOpencv. And then link libraries as you did with tensorrt in the previous step.
$ python -m pip install -r requirements.txtYou should expect performance for one processed image around the numbers in the table. I checked forums and some users said you can get better numbers if you will switch to INT8 precision and if you will use opencv3.8 on jetsons. If the numbers are not good enough check those ideas first. Probably I will try it out later if you are interested. But the numbers you see is as low as you can get with FP16.
Also, you can check runtime yourself, a message is published with debug flag.
| Model | GTX1080 | Xavier | nano |
|---|---|---|---|
| yolov3-tiny-288 | 10 ms | 25 ms | 80 ms |
| yolov3-tiny-416 | 13 ms | 30 ms | 135 ms |
| yolov3-416 | 33 ms | 95 ms | 660 ms |
| yolov3-608 | 62 ms | 195 ms | x ms |
I don't really care if you will reference me or will take the code without it. But from my side I think it is good to thank these people:
I referenced source code of NVIDIA/TensorRT. Those NVIDIA samples are under Apache License 2.0.
And thanks to jkjung-avt and his project with tensorrt samples. I took some parts from there.