Creating tensors from opencv data

[photex]

Howdy!

When trying to use this crate in combination with the opencv crate there appears to be no fast way to build a Tensor from an opencv matrix.

Currently you have to copy pixel by pixel and it absolutely kills my framerate. Without even running inference, just initializing the tensor this way takes me from ~22ms -> ~60ms of frame preparation time.

Before I trod down any paths you're currently looking into, I wanted to get your thoughts on approaches you'd accept a PR for.

What would make my life easiest would to be able to get a pointer to the tensor data, which I can construct a cv matrix from. I'd then use the cv api to initialize the tensor (normalizing the pixel values).

[photex]

I might have a working solution:

let mut input_tensor = tf::Tensor::<f32>::new(&[1, 512, 512, 3u64]);
let ptr = input_tensor.as_mut_ptr() as *mut c_void;
let mut input_mat = unsafe {
    Mat::new_rows_cols_with_data(
        512,
        512,
        cv::core::CV_32FC3,
        ptr,
        3 * std::mem::size_of::<f32>(),
    )?
};
cv::imgproc::cvt_color(
    &normalized,
    &mut input_mat,
    cv::imgproc::COLOR_BGR2RGB,
    0,
)?;

where normalized is an opencv matrix already converted to floating point and normalized to 0..1.