Optimized object detection algorithm issues

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
1

Hi all,
In my workplace we develop automotive application on the Jetson Xavier. We would like to run object detection algorithm as part of the system. We tried to convert some of the well known object detectors to TensorRT (for example YOLO v5).
The problem is that TensorRT doesn’t support all operators and therefore it’s performance degrades significantly.

Can you please recommend on high performance (both quality and run time) TensorRT based network for that purpose?

Thanks,
Guy.

3

Hi,

You can check the models recommended in the jetson-inference:

Thanks.

4

The models that appear in the link have lower performance (in respect to quality of detections) compared to YOLOv5, Retina and others.
Can you suggest an optimized version of the cutting edge detection algorithms of nowadays? Preferably in automotive field?

Thanks,
Guy.

5

Hi,

If performance is the major concern, you can check our TLT toolkit which can support pruning:

Thanks.

6

I mean they have lower performance in respect to the quality of the algorithm. Low rate of correct detections.
Thanks.

7

Hi,

Maybe you can check the database between YOLOv5 and jetson-inference.
If the database is relatively smaller, you can retrain it with your custom database for accuracy.

Thanks.

8

Hi @glifchitz ,
You can use YoloV4 with CSPDarkNet53 in Transfer Learning Toolkit or any of the following architectures and backbones and yield better accuracy-performance via QAT and Pruning

TLT3_2
TLT3_2954×386 17.8 KB

https://docs.nvidia.com/metropolis/TLT/tlt-user-guide/text/object_detection/yolo_v4.html