Person detection, split "loading layers of weight-file" and "image detection"

I want to use my Jetson Nano for person detection (in iobroker).

i recognized that actually the detection of the image takes less than a seconds. This more than fast for me.

My problem ist, that first part (Loaded 107 layers from weights-file)
./darknet detect cfg/yolov3.cfg yolov3.weights
takes arround 6 seconds

and then after i enter a image file i have the result within a second.

Is there a way that i don’t have always to do the first part? Can “the result of the first part” be stored somewhere?

I know that you recognize that i don’t have knowledge at all.

i would like to enter a linux shell command with a image and want the result in less tha 1-2 seconds. This wold be great

Means i want to do the first part (Loading layers from weights-file) once and then i want just to start the detection of images from the console (but not inside the “first part call”) --> “Enter Image path”

is there is really no way in order to split the loading from layers and the object detections?


It looks like you are using darknet interface.
On Jetson, it’s recommended to use TensorRT to get the optimal performance.
You can check our sample in Deepstream (use TensorRT) for YOLOv3.


Then, the sample can save the compiled TensorRT engine into file so you can de-serialize it next time.
This will save you the long model generation time from the second launch.


1 Like

I uses that link in order to download the tbz2 file:

The file is extracted.

I have now the directory and executed

@kayccc @AastaLLL
Can you give me a hint what i have to do next?

Even i am 20 years in IT business it is hard for me…i am almost giving up.

would be nice if you could help me what to do next. I don’t need many things, i just want to do object (persons) detection on a given image. Thats all.


Somehow the picture is black so I’m not able to know the issue you meet.

A common cause is the incorrect package are used.
For Jetson device, we recommend to install it directly from the sdkmanager.
You can find the Deepstream component in the STEP01.


See my post 5 of 7.
I went in the direction you recommended me. But i don’t know how to continue.
My impression is, that this NVIDIA “things” are only for experts or i just not good enough :-)

Or in other words, to be really even more successful you should improve documentations for beginners ;-)

I understood that the forum is nothing for totally beginners.
I have just one last question:
With darknet there is a solution for me that at least the second image i get the detection in less than a second.
The other way with TensorRT i went, but i was not able to get the detection less than 3-6 seconds with my jetson nano. For sure i did something wrong. What i have to find by myself, but now the question:

When i go the way TensorRT is there a way that i call the detection from the shell and get the detection result in less than 1 second if i do everything right? (with jetson nano). I have just this generell one question (and don’t ask how exactly this would work - i still miss a step-by-step manual for object detection - with every detail step description)

@AastaLLL thanks for (not) answering.
please delete to whole thread, it makes no sense.