How to access primary primary nvinfer object and user meta in nvdspreprocess?

[abramov_ov]

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) GPU
• DeepStream Version 6.0.1
• TensorRT Version 8.0.1.6
• Issue Type( questions, new requirements, bugs) Questions

My case is similar to the one described in How can I use nvdspreprocess before secondary nvinfer to do some perspective job?

Unfortunately, the post tagged as the solution in the linked question does not contain enough details.

Generally speaking, I’d like to implement the following pipeline:

<input, decode, nvstreammux> → primary nvinfer face bbox + landmarks detector → face objects preprocessing (some affine transform on each face using respective landmarks) → secondary nvinfer face classifier

Is it possible for nvdspreprocess element to apply preprocessing to detected objects, and not some predefined ROIs?

Is it possible for nvdspreprocess element preprocessing functions to get access to object meta (containg bboxes of detected objects) and user meta (containing landmarks tensors)?

If possible, please describe how to do it in more details.

[kayccc]

Sorry for the late response, our team will do the investigation and provide suggestions soon. Thanks

[Fiona.Chen]

nvdspreprocess is for pre-processing before inference. The object will only be generated after inference and post-process. There is no object available in nvdspreprocess.