GPU performance improvements

[wohali]

It would be nice if CUDA could be applied here.

The current state of affairs (September 2023):

• cuCIM accelerates scikit-image via GPU, but lacks any form of Hough line transform.
• OpenCV has a CUDA backend with the following limitations:
  • Manual movement of data between CPU and GPU is required; and
  • Only the classic Hough line transformation is provided, not the probabilistic implementation. This may be OK, but the output is in (θ,rθ) form rather than as start and end line endpoints.
  • The standard OpenCV library has cv2.HoughLinesP(), (sample code) and might be worth exploring on its own.
  • OpenCV also has a LineSegmentDetector (demo) (sample code) that may or may not be useful for our purposes.

[wohali]

A performance analysis of an average worker process shows most of the time is being spent in binary dilation (when peak sum Hough fails) and the median filter blur at the start of hough angle identification. By comparison actual probabilistic Hough edge detection is lightweight.

[wohali]

Notes to self on getting opencv with CUDA bindings working on Windows:

• find the latest wheel from: https://github.com/cudawarped/opencv-python-cuda-wheels/releases
• install corresponding CUDA Toolkit from https://developer.nvidia.com/cuda-downloads
• install appropriate cuDNN from same place - cut & paste contents into the install location for CUDA (under Program Files)
• optional install appropriate video codec from https://developer.nvidia.com/video-codec-sdk#Download
• test with import cv2