Yolov3 on Jetson AGX Xavier Developer Kit with ROS on a mobile robot

Good morning everyone,

i have to use yolov3 (specific and optimized) on a Jetson Xavier Developer Kit to detect (pedestrians and vehicles…) then return positions with ROS to central computer for fusion.

so i’m using Jetpack 4.1.1 and generic yolov3 runs at 22 fps in EDP/MAXN power mode (on coco data) to performance tests.

so my question is :

  • The software configuration that I chose is it optimal in this case? Feel the need for working with L4T adaptation or deepstream : knowing that the speed objective is limited by the lidar which runs at 15fps maximun (goal already reached) ?

Thanks.

Hi,

Do you run it with the darknet frameworks directly?

You can have a look on deepstream since it optimized the pipeline and remove all the non-necessary memory copy.
https://developer.nvidia.com/embedded/deepstream-on-jetson-downloads

Thanks.

Hi,

Sorry for my long response time, I had to work on demonstrations in the last days, and I still have more to do. But I keep a time later for optimization.

Yes I’m using darknet as framework. Actually using this repo (https://github.com/leggedrobotics/darknet_ros). So as you say, can deepstream help me to improve computation speed ?

Thanks.

Hi,

YES. We have optimized the pipeline and do some acceleration for it.

For YOLO, it’s recommended to check this tutorial first:
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/tree/master/yolo

Thanks.

Good morning,

Do you have trained the net again in order to achieve better detection performance of few classes (vehicles and pedestrians), or you are using the pretrained COCO weights (80 classes)?

Thank you in advance.

Good morning @vortizujbls,

Yes i have trained the net by my-self but not in order to achieve better detection performance but in order to detect custom classes.

Network has been trained by fine-tuning the darknet53.conv.74 weight file. The pretrained COCO weights are used to generate annotations and make it easy to create the database before fine-tune.

you’re welcome.

Ok,

We are having trouble to retrain the Yolov3 net to achieve better performance in detecting 6 existing classes only.

Do you use Darknet C environment to train?

Yes,

We fine-tune the darknet53.conv.73 weight file into Darknet environement given by this repo : https://github.com/AlexeyAB/darknet

I think that to improve the performance of yolov3 on existing classes must pass through the database, especially if you had a specific environment for your application.

Well, our intention is to train the net with specific on-road databases.

I advise you to use all that is already available for the classes that interest you, and to increase / complete this database with yours.

What do you think ?

Yes, that’s the objective (to use Caltech, KITTI… among others).