The documentation is not very specific regarding the question, what cluster-mode is doing what. Also, above what confidence it would make sense to deal with detections.
Is there some deeper insight?
Hi,
You may refer to sources/libs/nvdsinfer/nvdsinfer_context_impl_output_parsing.cpp, for cluster mode 0 CV group rectangle, see function DetectPostprocessor::clusterAndFillDetectionOutputCV, how it implemented, and we have comments, you can have one general understanding. you also can find other cluster mode related functions within this file.
please let me know if you have more questions.
Thanks. I was not after the current implementation, more after like a small description about the principles of each clustering mode. The more that I have reported a bug with them: Phantom detections in cluster modes != 0
Especially I’m curious about, why the cluster-mode 0 is unable to deliver confidence values. Why isn’t it possible to average the confidences of each of the clustered detections? Or even median it?
Sorry for the delay.
Did you mean the confidence value got was -1? if yes, check this,
gfloat confidence
Holds a confidence value for the object, set by the inference component.
confidence will be set to -0.1, if “Group Rectangles” mode of clustering is chosen since the algorithm does not preserve confidence values. Also, for objects found by tracker and not inference component, confidence will be set to -0.1
_NvDsObjectMeta — Deepstream Deepstream Version: 5.0 documentation
OK, thanks. Found that already. What value is a good “operational” value? Is this something to be found by trial and error, or is there a recommendation?
It’s not an error, the design is like this.
cv::groupRectangles accept type std::vector< Rect>, so m_PerClassCvRectList does not include confidence info
check DetectPostprocessor::clusterAndFillDetectionOutputCV
You got me wrong. The question was: What would be a recommended value for this? I mean, 0.2 is obviously. not that good, since every box is detected as car.
But aren’t you asking about confidence value when cluster mode set to 0 CV grouping?
No. I understand, that you are making a difference here.
i thought you asking about this?
i am confused.
You got another question, right?
below question?
What would be a recommended value for this? I mean, 0.2 is obviously. not that good, since every box is detected as car.
Please disregard