use camera with yolo and deepstream

[elementdl]

Hi @marcoslucianops ,

I want to test Yolov3-tiny with a camera plug and play. In deepstream_app_config I changed the type to 1 and I got some errors, it says it failed to creat_camera_source_bin.

Do you know how to use yolov3-tiny with a simple usb camera?

For now I just started to use deepstream, by any way do you know if is there some tutorial to begin with deepstream?

Sincerely,

[marcoslucianops]

Do you know how to use yolov3-tiny with a simple usb camera?

Try to use usb camera changing [source0] lines to

[source0]
enable=1
type=1
# Resolution
camera-width=1280
camera-height=720
# FPS
camera-fps-n=30
camera-fps-d=1
# /dev/video0
camera-v4l2-dev-node=0

For now I just started to use deepstream, by any way do you know if is there some tutorial to begin with deepstream?

I don't know any tutorials. To use deepstream according to your need, you need to edit C/C++ or Python deepstream app code (deepstream-test1, deepstream-test2, etc.) or edit gst-dsexample (C/C++).

[elementdl]

Thanks a lot it works!

For example how did you know that I had to add these modification?
Is there a code somewhere who explain how to manipulate deepstream like this? (in deepstream-test1, deepstream-test2,...?)

Also, it is true that the programm is working really fast, but with Yolov3-tiny not modified it seems that it doesn't work so well : I tried to show myself with a webcam and the bounding boxes that recognize a person don't stay on the display like the program don't really recognize me. Is it normal? I followed all your steps on github, I just add the weight of yolotinyv3 directly from darknet.

Also i'm a begginer in computer vision and object recognition. What would you recommand for a begginer to practice and understand this domain?

[marcoslucianops]

For example how did you know that I had to add these modification?

I was seeing several topics in the NVIDIA forum to learn.

Is there a code somewhere who explain how to manipulate deepstream like this? (in deepstream-test1, deepstream-test2,...?)

You need manipulate tiler_src_pad_buffer_probe function to do your need. This function receives Frame Meta, Object Meta, etc.

Also, it is true that the programm is working really fast, but with Yolov3-tiny not modified it seems that it doesn't work so well : I tried to show myself with a webcam and the bounding boxes that recognize a person don't stay on the display like the program don't really recognize me. Is it normal? I followed all your steps on github, I just add the weight of yolotinyv3 directly from darknet.

Try to edit nvdsinfer_custom_impl_Yolo/nvdsparsebbox_Yolo.cpp (lines 376-377) from:

        //{0, 1, 2}}; // as per output result, select {1,2,3}
        {1, 2, 3}};

To:

        {0, 1, 2}};

And see if it makes any difference.

In my custom model, I have good accuracy, but I haven't properly tested the yolov3-tiny darknet model.

Also i'm a begginer in computer vision and object recognition. What would you recommand for a begginer to practice and understand this domain?

I'm a beginner in computer vision too. I recommend that you see the NVIDIA forum and docs to learn more.

(Sorry for any English error, it's not my native language)