RGB 24-bit support

odedoded
in KL720 related discussion

Hi,

We have a model that was trained on 24bit RGB images (ssd mobilenet 3x320x320). In the Kneron SDK we can only see support for conversion to RGB565 and RGBA8888. Using RGBA8888 performs better than RGB565, but it's still less than what was achieved during training.

Is there any option to use regular 24bit RGB?

The default img_format (part of the model descriptor) is ImageFormat.KP_IMAGE_FORMAT_RGBA8888 but the model has only 3 channels (channel: 3, height 320). Is the image converted to 24 bit somewhere? is there a way to set these parameters during the toolchain conversion?

Thanks!

Comments

  • Maria ChenMaria Chen

    Hi Oded,

    No worries, there's no need to convert anything. You can use 24bit model as well. When it is being converted using the toolchain, it will automatically fit the NPU format (RGBA8888).


    You will need to input the right parameters for pre-processing: KP_IMAGE_FORMAT_RGBA8888. Please make sure your image is converted into the right format.

    Below is an image of how the image gets sent, with R first, then G, then B, then A. In your case, since A wouldn't be used, you could set the A to 0.

    image.png


The discussion has been closed due to inactivity. To continue with the topic, please feel free to post a new discussion.

© Kneron, Inc.