This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AM62A3: Processors forum

Part Number: AM62A3

Dear TI Expert,

We have an image processing algorithm model in ONNX format. The input to the model consists of four RGGB channels. During the model importation using the TIDL-RT tool on the AM62A, we have set the `inFileFormat` to 2, and the `inData` is provided in a binary format. Since our input is a raw image rather than RGB, BGR, or YUV, how should we set the `inDataFormat` parameter? Can the TIDL-RT import tool successfully handle the importation of an algorithm model with four RGGB channels?

  • Hello Tian Liang,

    I am not sure that we support this input format, since it is assumed that the RGGB images from the image sensor will pass through the ISP to produce a fully chromatic RGB image. I am figuring this out in the background. In the meantime:

    Can you help me understand your system a bit better? The raw image sensor will produce a Bayer-mosaic image where each pixel is a different color (R G or B, depending on the pixel position within the repeating mosaic pattern). We consider that Bayer image to be RGGB (or some other series of letters describing the pattern). This is technically a single channel image of the sensor's resolution, and will become multiple channels (resolution) once the ISP interpolates between same-colors within the Bayer/RGGB image.

    Are you saying that you are using 4 of these Bayer/RGGB images, stacked into multiple channels, or you are rearranging the RGGB image into multiple channels?

    My suggestion would be to use the Python OSRT interface to compile this model and handle the input preprocessing in your own application code before passing images to the ONNX interface. The TIDL-RT import interface does not indicate that it has support for reading and preprocessing raw images like these.

    Best,
    Reese