When using CUDA 12.6, what is best way to perform gaussian blur in python?

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
,
1

I understand there are at least 2 ways to do this:

  1. Using vpi module
  2. Using CV2

I’ve tried both,
1.

(python_env_01) nano@nano:~/Documents/yolo_stuff$ python3
Python 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import vpi
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'vpi'
>>> 
[1]+  Stopped                 python3
(python_env_01) nano@nano:~/Documents/yolo_stuff$

nano@nano:~$ python3
Python 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import vpi
>>>

‘find /usr/lib/python3 -name vpi.so 2>/dev/null’ yields nothing

I want the python environment to remain closed from the system installed stuff to avoid conflicts, but to allow module vpi if possible?

Python code snippet:

   # 1. Create a GpuMat object and upload the CPU frame to the GPU
    # GpuMat objects hold data in GPU memory
    gpu_frame = cv2.cuda.GpuMat()
    gpu_frame.upload(frame)

    # 1. **FIXED:** Define the correct base type integer (depth)
    # DIAGNOSTIC RESULT: Input was 'uint8' (Depth 0), which corresponds to cv2.CV_8U.
    GPU_DEPTH_TYPE_INT = cv2.CV_8U 

    # 2. Apply the Gaussian Blur using the CUDA module
    # We must use the createGaussianFilter factory function and the .apply() method.
    
    # Create the reusable Gaussian filter object
    # Arguments: 1. srcType, 2. ksize, 3. sigmaX, 4. sigmaY, 5. dstType
    cuda_filter = cv2.cuda.createGaussianFilter(
        GPU_DEPTH_TYPE_INT,         # 1. srcType (Using cv2.CV_8U)
        (31, 31),                   # 2. ksize
        0,                          # 3. sigmaX
        0,                          # 4. sigmaY 
        GPU_DEPTH_TYPE_INT          # 5. dstType (Using cv2.CV_8U)
    )

    # Apply the filter to the GPU frame
    gpu_blurred_img = cuda_filter.apply(gpu_frame)
    
    # 3. Download the result back to a NumPy array in CPU memory
    blurred_img = gpu_blurred_img.download()

Result:


(python_env_01) nano@nano:~/Documents/yolo_stuff$ cd && python3 -m venv python_env_01 && source python_env_01/bin/activate && export LD_LIBRARY_PATH=/home/nano/python_env_01/lib/python3.10/site-packages/nvidia/cusparselt/lib:$LD_LIBRARY_PATH && cd && cd /home/nano/Documents/yolo_stuff && python3 test_yolo_on_video_file_onions_CUDA_17.py
CUDA is available for cv2.
Load a model ....
Starting object detection on video file: /media/nano/KIOXIA/Videos/best/my_webcam_recording_SAT_WE_271111_fixed_shortened.mp4
2025-10-02 15:54:31.984123028 [W:onnxruntime:Default, device_discovery.cc:164 DiscoverDevicesForPlatform] GPU device discovery failed: device_discovery.cc:89 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor"

[DEBUG] Input NumPy Frame Data Type (dtype): uint8
[DEBUG] GpuMat Depth (gpu_frame.depth()): 0
Traceback (most recent call last):
  File "/home/nano/Documents/yolo_stuff/test_yolo_on_video_file_onions_CUDA_17.py", line 323, in <module>
    frame_with_all_drawings, green_centers = process_frame_for_green_centers(
  File "/home/nano/Documents/yolo_stuff/test_yolo_on_video_file_onions_CUDA_17.py", line 172, in process_frame_for_green_centers
    cuda_filter = cv2.cuda.createGaussianFilter(
TypeError: Argument 'dstType' is required to be an integer

what is the correct code layout i should be using - are there any examples for cuda 12.6?

Please do not reply using LLM.
Thank you.

5

Hi,

It’s recommended to use VPI since we have optimized it for the Jetson device.
You can test a blurring sample with below steps:

$ /opt/nvidia/vpi3/bin/vpi_install_samples.sh .
$ cd NVIDIA_VPI-3.2-samples/tutorial_blur/
$ python3 main.py ../assets/kodim08.png

The output tutorial_blurred_python.png can be found in the tutorial_blur folder.
The sample demonstrates with a box filter. Please refer to the link below to change it to Gaussian: (switch Language to ‘python’)

Thanks.