Darknet: Deepstream YOLOv3-tiny

[elementdl]

Hi @AlexeyAB,

I tried Deepstream several times and I don't get good detection results.
If I compare deepstream and darknet on the Jetson Nano, with the same .cfg and .weights, darknet run slower (because there is no utilisation of tensorRT) but the detection is good, we can see boundings boxes where there are objects.
Deepstream run way faster but detect almost nothing.

Is it normal? Is there a way to understand why it doesn't work?

Sincerely,

[AlexeyAB]

• Did you test default yolov3.cfg/weights model?

• If you use your own model, try to train your custom model with [net] letter_box=1 in cfg. And the run detection in Darknet with -letter_box flag the the end of detection command. And use this model with TRT.

• Did you check that you use TRT correctly?

[marcoslucianops]

@AlexeyAB

DeepStream uses NMS_THRESH = 0.3f and PROB_THRESH = 0.7f. What's the NMS_THRESH and PROB_THRESH used in Darknet demo?

[AlexeyAB]

@marcoslucianops

nms = 0.45 https://github.com/AlexeyAB/darknet/blob/1fb4606d5d04a0438887ca9a5681af25291fd240/src/demo.c#L248

---

conf_thresh = 0.25 https://github.com/AlexeyAB/darknet/blob/1fb4606d5d04a0438887ca9a5681af25291fd240/src/detector.c#L1878

[marcoslucianops]

nms = 0.45
conf_thresh = 0.25

With this values the detection is more similar to darknet

@elementdl, you can change lines 307-108 in nvdsparsebbox_Yolo.cpp

    static const float kNMS_THRESH = 0.3f;
    static const float kPROB_THRESH = 0.7f;

to

    static const float kNMS_THRESH = 0.45;
    static const float kPROB_THRESH = 0.25;

And see if you get better results.

[marcoslucianops]

My results

NVIDIA DeepStream SDK: https://youtu.be/a0yRw19Jb78
Darknet: https://youtu.be/rAJumTq17Eo

Edit: Used mask = 0,1,2 in both videos.

[AlexeyAB]

For default weights use
mask = 1,2,3
instead of https://github.com/AlexeyAB/darknet/blob/1fb4606d5d04a0438887ca9a5681af25291fd240/cfg/yolov3-tiny.cfg#L175

[daliel]

I’m not sure, but it’s possible that the preprocessing of a image in a deepstream affects the accuracy.
The pre-processing function is:
y = net-scale-factor * (x - mean)
Where:
•x is the input pixel value. It is an int8 with range [0,255].
•mean is the corresponding mean value, read either from the mean file or as offsets[c], where c is the channel to which the input pixel belongs, and offsets is the array specified in the configuration file. It is a float.
•net-scale-factor is the pixel scaling factor specified in the configuration file. It is a float.
•y is the corresponding output pixel value. It is a float.

[AlexeyAB]

For Yolo and the most modern models scale=1 mean=0. So Darknet doesn't use this.

•net-scale-factor is the pixel scaling factor specified in the configuration file. It is a float.

Can you show this value?

May be this is related to quantization?

Did you set mask = 1,2,3 in the last [yolo] layer?

[daliel]

Can you show this value?

net-scale-factor=1

May be this is related to quantization?

network-mode=0
some small boxes are different from darknet
What mask use for yolov3(full)?

[AlexeyAB]

• For yolov3-tiny.cfg - use mask = 1,2,3 in the last [yolo] layer
• For all other models - use default masks

---

y = net-scale-factor * (x - mean)

Try to use y=x instead of this line and look at the result

[daliel]

Sorry, I couldn’t find the code you want. I created a men-file filled with zeros. In accordance with the formula, it turned out:

y = net-scale-factor * (x - mean)
where:
    net-sacale-factor = 1
    mean = 0
then:
y = 1* (x-0) = x

The situation has not changed

[AlexeyAB]

Do you use yolov3-tiny.cfg?
Did you set mask = 1,2,3 in the last [yolo] layer?

[marcoslucianops]

NVIDIA DeepStream SDK: https://youtu.be/zDwisWLCM08

YoloV3-Tiny

NMS_THRESH = 0.45;
PROB_THRESH = 0.25;

mask = 1,2,3

net-scale-factor=1
network-mode=0 (FP32)

[elementdl]

hi,

Thank you all for your response.

@marcoslucianops I tried to apply theses modifications, but I sadly didn't get any better results, In your videos the bounding boxes stays consistent or detectect way more object, in my video it's the opposite.
You used a jetson nano too, right? At maximum power?

I think I will try to reinstall deepstream and try again, if it's working for you, I don't see why it would be not working for me.

[marcoslucianops]

You used a jetson nano too, right? At maximum power?

Yes for both.

Test with these files: yolov3-tiny.zip
Only add yolov3-tiny.weights.
You need make before test (CUDA_VER=10.0 make -C nvdsinfer_custom_impl_Yolo)

[elementdl]

Ok, I don't know what is missing in my code but yours works. I can't thank you enough for your help! I've been stuck with this issues for 3 weeks.