Invalid GPU Ordinal in function 'NvidiaOpticalFlowImpl'

Gaming and Visualization Technologies Video Codec, PyNv & OFA
1

Hi there,

I want to develop a optical flow pipeline using opencv and gpu support (RTX 2070)

I installed opencv 4.2.0 from source and also the contrib modules.
I installed the optical flow sdk.

When I run the code from:

https://devblogs.nvidia.com/opencv-optical-flow-algorithms-with-nvidia-turing-gpus/

I end up getting the error:

python3 ofgpu.py 
Traceback (most recent call last):
  File "ofgpu.py", line 12, in <module>
    nvof = cv2.cuda_NvidiaOpticalFlow_1_0.create(frame1.shape[1], frame1.shape[0], 5, False, False, False, 1)
cv2.error: OpenCV(4.2.0) /home/user/opencv_contrib-4.2.0/modules/cudaoptflow/src/nvidiaOpticalFlow.cpp:321: error: (-5:Bad argument) Invalid GPU Ordinal in function 'NvidiaOpticalFlowImpl'

How can I solve this?

I can run the two samples from the optical flow sdk on my gpu without a problem with cuda 10.0

2

Hi.

nvof = cv2.cuda_NvidiaOpticalFlow_1_0.create(frame1.shape[1], frame1.shape[0], 5, False, False, False, 1)

In the above line, last argument is gpu ordinal. If you have 1 gpu in the system, then gpu ordinal should be 0. Implementation of this API tries to create a CUDA context based on the gpu ordinal passed. If such a device is not present physically, this error occurs.
Please try passing 0 as gpu ordinal instead of 1.

Thanks.

3

Thanks a lot,

is there any information about the python nvof opencv functions, or where is one supposed to find this information?

4

Hi.
You can refer to OpenCV’s auto generated documentation based on the doxygen tags in the source code.
Please refer
OpenCV: cv::cuda::NvidiaOpticalFlow_1_0 Class Reference
OpenCV: cv::cuda::NvidiaHWOpticalFlow Class Reference

Thanks.