Image Transformation (crop) for one specific secondary classifier

Setup Specifications:

• Hardware Platform Quadro P4200
• DeepStream 5.0
• TensorRT 7.0.0-1+cuda10.2
• NVIDIA GPU 450.66


I am using a primary detector to detect different classes and two secondary detectors to recognize other attributes.

I would like to input to only one of the secondary detectors, a specific part of the detected objects e.g. central part of an object of type Vehicle (in other words: Feed a cropped image of the object that was output by the primary detector).

I have read in another topic that secondary detectors are nvinfer plugins and they crop images according to metadata (bounding boxes coordinates) from the primary detector.

Would you tell me please how to achieve this for only one secondary detector and maintain the default behavior for the other secondary detector?

Thank you in advance.

Can you try with process-mode=2?

The Gst-nvinfer plugin can work in two modes:
•Primary mode: Operates on full frames
•Secondary mode: Operates on objects added in the meta by upstream components

Is this still an issue to support? Any result can be shared?


thank you for your response.

Using the secondary mode only didn’t satisfy my use case.

In addition to the usage of the secondary mode, I would like to cut/crop the objects according to some static value (for example: Cut the object vertically into two equal parts) and then feed the result to a set of secondary detectors.

Does this can meet your requirement?



Thank you for your response.

Are these parameters that one needs to define/specify in the configuration file of the secondary classifier?

And where/how can I configure the width and height of the cropping?

Thank you in advance.

please check usage of nvvideoconvert
gst-inspect-1.0 nvvideoconvert
you could do it in config, or you could specify the crop property directly in your code.