VPI library : request for new functions


i have started to use VPI with Jetson AGX Orin. This library is really cool and performs great.

For information, my work :

I really hope you will add new functions in VPI.

Here are my requests :

Color conversion : add the possibility to debayer bayer image (4 main bayer patterns at least)

Color image channel split and merge like Python cv2.split() and cv2.merge(). For now, VPI channel mixer returns 3 color images (R, G, B). It would be better to get R, G, B mono channels

It would be very useful to get opencv equivalent functions like cv2.subtract, cv2.threshold, cv2.contrastCLAHE, cv2.addWeight and so on for VPI (only mono channel functions).

It would also be really useful to get more statistic functions like percentile for example (always mono channel functions).

I hope VPI will get further versions because it is very very good library.

It would also be really cool if we could get a VPI version for Windows !

Ok to discuss with Nvidia about those demands.


Is there someone from NVidia here ?

I have a simple problem :

I want to perform gaussian filter on a numpy array monochrome image and get back the result in a numpy array.

Here is my code :

    backend = vpi.Backend.CUDA
    image_vpi = vpi.asimage(image_mono)
    with backend:
        image_vpi = image_vpi.convert(vpi.Format.U8)
    with backend:
        gaussian = image_vpi.gaussian_filter(niveau_blur, 1.7, border=vpi.Border.ZERO)
    PIL_Image = PIL.Image.fromarray(gaussian.cpu())
    image_gaussian_blur_mono = np.array(PIL_Image)

I get this error message :

PIL_Image = PIL.Image.fromarray(gaussian.cpu())
RuntimeError: VPI_ERROR_INTERNAL: Can’t perform shared mapping

I always have this issue, whatever i do (i tried many things.

What is the solution ?


Well, it seems NVidia does not no the existence of Computer Vision & Image Processing posts !

Still wait for an answer.