Deepstream v4: zero confidence problem

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

I’m using “/deepstream-4.0/sources/objectDetector_Yolo/nvdsinfer_custom_impl_Yolo/nvdsparsebbox_Yolo.cpp” as a custom parser. My problem is in “attach_metadata_detector” function in file : “gst-nvinfer/gstnvinfer_meta_utils.cpp” #Ln83 :
obj_meta->confidence = 0.0;

How can I change this to get the confidence calculated before in “nvdsparsebbox_Yolo.cpp” ??
Is there any way to access NvDsInferObjectDetectionInfo object from “attach_metadata_detector” function ??

Thanks.

2

Get confidence here:

nvdsinfer_context_impl_output_parsing.cpp

parseBoundingBox() {
    ...
objectList.push_back({ classIndex, (unsigned int) rectX1,
                        (unsigned int) rectY1, (unsigned int) (rectX2 - rectX1),
                        (unsigned int) (rectY2 - rectY1), confidence});
...
}

I think you can add “confidence” parameters in “NvDsInferObject”
and set the value in “NvDsInferContextImpl::clusterAndFillDetectionOutputCV”
and then it can get confidence in “attach_metadata_detector()”

3

Thanks a lot. I’ll try it and let you know how it goes.

4

I couldn’t make it work. I’ve added “confidence” parameters in “NvDsInferObject” and modified “NvDsInferContextImpl::clusterAndFillDetectionOutputCV” :

for (unsigned int c = 0; c < m_NumDetectedClasses; c++)
    {

        /* Add coordinates and class ID and the label of all objects
         * detected in the frame to the frame output. */
        for (auto & rect:m_PerClassCvRectList[c])
        {
            NvDsInferObject &object = output.objects[output.numObjects];
            object.left = rect.x;
            object.top = rect.y;
            object.width = rect.width;
            object.height = rect.height;
	    object.confidence = 0.5;
            object.classIndex = c;
            object.label = nullptr;
            if (c < m_Labels.size() && m_Labels[c].size() > 0)
                object.label = strdup(m_Labels[c][0].c_str());
            output.numObjects++;
        }
    }

Also, I modified “attach_metadata_detector()” :

obj_meta->confidence = obj.confidence;

I rebuild “nvdsinfer” and “gst-nvinfer” libraries.

However, I get :
confidence = 4747328616950188978942771200.000000

This doesn’t make any sense.

Did I missed something ??

5

I tried to set the values of object.left, object.top, object.width, object.height to zero in “NvDsInferContextImpl::clusterAndFillDetectionOutputCV” function.
I rebuild everything.
I still got non null bbox detections (which doesn’t suppose to happen).
This means that “NvDsInferContextImpl::clusterAndFillDetectionOutputCV” doesn’t do anything.
I even tried modifying “NvDsInferContextImpl::clusterAndFillDetectionDBSCAN” function and still got the same results.

I’m out of ammo on this one.

I’d appreciate any help possible.

Edir : It finally worked. Anyone having the same problem refer to : [url]https://devtalk.nvidia.com/default/topic/1058661/deepstream-sdk/nvinfer-is-not-populating-confidence-field-in-nvdsobjectmeta-ds-4-0-/post/5373361/?offset=13#5373383[/url].

6

attach nvinfer process here. DBScan have confidence, but not for opencv. I think you can try to add confidence as my comment 2. We will fix this issue in the next version.

External Media

7

Hi ChrisDing,

is it possible to fix confidence problem without modifying the SDK after every new version release??

Thanks

8

I checked 4.0.2 code, confidence has not been added.
5.0 will add it.

9

Hi ChrisDing,

Do you know when the 5.0 version will be realeased ?

Thanks

10

5.0 is under development and will be announced in the next few months.

11

Even with 5.0 there is no useful confidence value.

1 Like