search

dakotabenjamin / CameraCalibration

GNU General Public License v3.0
33 stars 19 forks source link

Undistort performance #12

Closed sberryman closed 6 years ago

sberryman commented 6 years ago

I ran into this issue on a side project where I had to undistort ~7 million 4K resolution images. It turns out OpenCV's undistort function computes a rectification map every execution.

If you simply move getOptimalNewCameraMatrix before the for loop, add mapx, mapy = cv2.initUndistortRectifyMap(K, d, None, newcamera, (w, h), 5) below that and then replace line 83 cv2.undistort(...) with cv2.remap(img, mapx, mapy, cv2.INTER_LINEAR) you will see a massive speed up. The only issue is that every image must have the same dimensions.

https://github.com/dakotabenjamin/CameraCalibration/blob/708694bd42ff18e204beb142925b9ccdecc83884/undistort.py#L82-L83

dakotabenjamin commented 6 years ago

Great find! I'll have to include a case where not all the images have the same dims but I'll check this out.

dakotabenjamin commented 6 years ago

Added.

sberryman commented 6 years ago

Nice! Did you test it on a large batch of images to see the performance improvement?

On a side note I've got ODM running on 1,443 images with some changes to depthmap parameters to depthmap_resolution=1280, depthmap_min_patch_sd=2.5 and depthmap_method=BRUTE_FORCE (Computer has an overclocked Intel Core i7-7820X w/ 64 GB ram)

The mapping area consists of an EXTREMELY dense urban environment. Thus, I'm not expecting amazing results. It has been running since 2018-03-02T21:47:40Z and moved to odm_25dmeshing roughly 8 hours ago.