Please provide complete information as applicable to your setup.
**• Hardware Platform : GPU
**• DeepStream Version : 6.1
*• Issue Type: questions
• Requirement details( This is for new requirement. Including the module name-for which plugin or for which sample application, the function description)
I saw the sample code ‘deepstream-3d-action-recognition’, and the description of nvinfer on the official website, it seems that ‘input-tensor-from-meta’ is applicable to PGIE only .
Is it possible to use custom input tensor metadata (input-tensor-from-meta=1) as input of SGIE ? I don’t find any restriction about this parameter in the code of nvinfer. Could someone help me ?
DeepStream 6.1 version does nor support (input-tensor-from-meta=1) with SGIE. Plesae wait for future versions or you can implement by yourself since gst-nvinfer and gst-nvdspreprocess are all open source.
Thanks for your reply. Could you tell me why there is such a restriction? and where is the source code about this restricted? cause i tried to find the code, but failed.
nvdsperprocess is still alpha version now The new features are on-going. Gst-nvdspreprocess (Alpha) — DeepStream 6.1.1 Release documentation
I knew the plugin, and have tried to read it’s code too. but I didn’t find the exactly codes that restricts ‘input-tensor-from-meta=1’ to PGIE. Could you help me?
Currently the batched tensor meta data is attached to batch meta directly, there is no object meta level tensor meta supported. So the object meta level tensor meta should be added for SGIE case.
Please investigate nvinfer first, you must understand how PGIE and SGIE works. And then read the nvdspreprocess code to understand how preprocessing is moved outside nvinfer when input-tensor-from-meta=1. There is no shortcut to understand the whole mechanism.
OK, thank you very much for your patient answer，it helps me a lot.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.