Tensorrt_yolo issue in autoware universe

[amtommy]

Hello!
I tried to run the tensorrt_yolo, but I got the incorrect detection results. (see figure)
How to fix it?

Thank you!

[shmpwk]

You can customize your own config parameter here
such as iou_thresh, ignore_thresh, or anchor size.

[amtommy]

Thanks for your reply.

I changed the threshold, but I still got the incorrect result. [figure]

• score_thread = 0.5
• iou_thresh = 0.5
• ignore_thresh = 0.8
• Onnx file from autoware.universe [link]

I used the Yolo official image (dog.jpg) to recognize, but also got the incorrect result. [figure]

[shmpwk]

Hmmm... weird...but as you mentioned the link, the network weight and conf is based on the open model for example, and we fine-tune the network to adapt our domain. You can also fine tune the model if you have your own dataset.

If you would like to get more precise results for your custom image, I recommend to use the upper version of yolo or other SOTA model.
If you would like to use image recognition for Autoware, you can avoid the problem for now since currently we use yolo mainly for the traffic light recognition here. As far as other objects recognition like vehicle, Autoware uses the combination of lidar points and image recognition here. We are trying to imporove the accuracy of the image recognition day by day.

[mitsudome-r]

@amtommy Could you provide a rosbag + launch file that you used so that other engineers can look for the cause?

[amtommy]

Details are provided below:

• rosbag: The topic name is /cam/front_bottom_60.
• launch file
• original parameter
• original model

[xmfcx]

Let's solve it in this issue: autowarefoundation/autoware.universe#918