Orb Cuda computeAsync not using updated keypoints

[IliyaPru]

I am using the ORB cuda CV2 detector. I need to filter the keypoint returned from detect and then compute the descriptors using the updated keypoints.

Unfortunately it seems that the computeAsync uses the keypoints received from the detect.
The number of keypoints returned from detectAsync is 3000 (max number defined in orb cuda create). The filtered keypoints is 1700 but the number of descriptors is 3000.

I know that I can use detectAndComputeAsync and filter after but is much time consuming.

I am using opencV 4.8.0 with python 3.8 on windows 64bit PC.
thanks,
Iliya

Code example
##################################
#detect keypoints
cuMat1 = cv2.cuda_GpuMat()
cuMat1.upload(image)
keypoints_detected_GPU = detector.detectAsync(cuMat1, mask=None)
keypoints_detected = detector.convert(keypoints_detected_GPU)

        #filter keypoints
        keypoints = FilterFunction(keypoints_detected )
        
        #upload filtered keypoints
        keypoints_np = np.array([
            [kp.pt[0] for kp in keypoints],  # x-coordinate
            [kp.pt[1] for kp in keypoints],  # y-coordinate
            [kp.response for kp in keypoints],  # response
            [kp.angle for kp in keypoints],  # angle
            [kp.octave for kp in keypoints],  # octave
            [kp.size for kp in keypoints],  # size
        ], dtype=np.float32)
        
        keypointsGpu = cv2.cuda_GpuMat()
        keypointsGpu.upload(keypoints_np)

        #Compute descriptor
        keypointsGpuCompute, descriptorsGpuCompute = detector.computeAsync(cuMat1, keypointsGpu)
        keypoints_np1 = keypointsGpuCompute.download()
        
        keypoints_list = []
        num_keypoints = keypoints_np1.shape[1]
        
        for i in range(num_keypoints):
            x = keypoints_np1[0, i]
            y = keypoints_np1[1, i]
            response = keypoints_np1[2, i]
            angle = keypoints_np1[3, i]
            octave = keypoints_np1[4, i]
            size = keypoints_np1[5, i]
        
            # Create cv2.KeyPoint object and append to the list
            kp = cv2.KeyPoint(x, y, size, angle, response, int(octave), -1)
            keypoints_list.append(kp)
        keypoints = keypoints_list

        #download descriptors
        descriptors = descriptorsGpuCompute.download()

[outoftolerance]

This issue appears to be present in the underlying OpenCV CUDA ORB implementation and not only in the Python binding. There is an issue open for it on the OpenCV repo as well (opencv/opencv#9151) that was opened in 2017.