Nvinfer and 16bit/color channel raw image data?

I have a TensorRT segmentation model that expects RAW 16bit/color channel input data and has been trained on such data.
I saw that the nvinfer plugin expects input in NV12 format.
Is there a way to perform inference on raw data using nvinfer? If not, is there any way to extend the functionality to perform inference on raw data myself? As I understand nvinfer is not open source.

  1. could you share your setup with us?
  2. nvinfer is open sourced and nvinfer can accept RGB/BGR and gray image.
  3. what’s “16bit/color channel input data”, could you explain more

Thanks for your reply!
Here is more information regarding your questions:

  1. I have a GStreamer camera source with raw images and run deepstream on a AGX Xavier.
  2. Can you provide a link to the source code of nvinfer then? I have trouble finding it.
  3. It’s 16bit/color channel raw image data, which results in a 48bit image data (for example TIFF, PNG 16bit formats). Normally neural networks are not trained on raw data but I trained a custom network for some experiments and it works flawlessly performing inference on my TRT engine directly.

edit regarding 2: I found the source code and confirmed your comment regarding RGB/BGR and gray image support. Sadly it seems to only support RGB with 8bit/pixel color depth and not 16bit. I saw GStreamer supports RGB 16bit/pixel data, is there any option to add support?

Hey, will check internally and give you a reply later?

Thank you!