Write and read data into pipeline

[erence]

Please provide complete information as applicable to your setup.

• Jetson Orin NX
• DeepStream Version 6.3
• JetPack Version 5.1
I am working on a GStreamer pipeline that processes video frames for object detection. My input frames have a resolution of 1920x1200. To improve recognition accuracy, I need to add padding to these frames, enlarging them to a resolution of 2200x2000. However, the model I’m using with nvinfer is optimized for smaller input sizes, which introduces complications.

Here’s my intended pipeline strategy:

1. Increase the frame size by adding padding to reach a resolution of 2200x2000.
2. Use a tee element to create two branches.

• Branch A: Scale down the resolution for compatibility with nvinfer.
• Branch B: Maintain the padded frame for further processing. I

3. Perform object detection on Branch A, then apply the detection coordinates of Branch A to crop the original, padded frame in Branch B (the coordinates will be scaled too).

My challenge is transferring the detection coordinates from Branch A to Branch B without causing frame desynchronization. Specifically, I need a reliable method to ensure that the coordinates extracted from the scaled-down frames in Branch A are applied to the corresponding, original frames in Branch B for cropping purposes.

Could you provide guidance on achieving this synchronization? Are there best practices or specific GStreamer elements/plugins that facilitate this kind of data transfer between branches specifically between nvidia gst plugins, ensuring that the correct metadata is applied to the appropriate frame?

Thanks for your help…

[yuweiw]

Could you try to add a nvvideoconvert between the 2 nvinfers?
—>nvinfer->nvvideoconvert(scale to 2200x2000)->nvinfer->…

[erence]

How can I put nvvideoconvert between two seperate pipelines?

If you mean on the same pipeline, it goes first to lower resolution. If I scale it up again, do I not lose data? Or is it a real 2200 resolution again in a gst pipeline??

[yuweiw]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

In fact, the actual scenario of our pipeline is as follows. The processing of images in nvinfer does not affect the images in the pipeline.

...nvstreammux(2200x2000)->pgie(2200x2000)->sgie(2200x2000)->...

The image processing in nvinfer is a in-place algorithm, it won’t affect the other nvinfers.