Yolov3 image detector very slow

Hello every one,
I’m using a Jetson Nano 4GB with Yolov3 to apply the image detector in live video (I have a webcam - USB camera - Logitech camera).
When I run the command:
./darknet detector demo cfg / coco.data cfg / yolov3.cfg yolov3.weights

The camera works well, but it’s very slow, I only get 1.5 fps.

I tried to change the dimensions of the camera in Yolov3.cfg to (width=416, height=416)

but the live video is always (1280*960) very slow, but this change was already applied to the test photos when I run the command:
./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg.

how can I adjust the dimensions of the camera in live video, or should I do something different to get more fps for live video?

I use Jetson nano 4GB
JetPack: 4.5
Ubuntu: 18.04
Python: 3.6.9
Torch: 1.8.0
torchvision: v.0.9.0
OpenCV: ‘3.2.0’

can you help me please?


YOLO is a complicated model so we don’t expect you can get the real-time fps on a source-limited platform like Nano.

Is YOLOv3 Tiny and Deepstream an option for you?
If yes, please check the below topic below which should reach 20fps.


Hi @AastaLLL ,
That is to say, I must first install the DeepStream SDK, NVIDIA SDK Manager and then make changes to the cfg files, or?



Please give it a try. The library is expected to have better performance.


That frame rate is about right for a nano running a full model yolov3 with that image size. From memory, in the past with the original PJ Reddie code is was around 600ms for inference with your settings.

Note, on the newer nano, the Orin, the frame rate of yolov7, a much better model, is just over 7 fps. But the memory required for yolov7 (pytorch based) needs the 8GB model.

The quality of recognition drops significantly with the tiny versions, for people detection I consider them unusable because of a very high false positive rate.


This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.