I am currently using OpenCV’s built-in patch-based histogram back projection (cv::calcBackProjectPatch()) to identify regions of a target material in an image. With an image resolution of 640 x 480 and a window size of 10 x 10, processing a single image requires ~1200 ms. While the results are great, this far too slow for a real-time application (which should have a processing time of no more than ~100 ms).
I have already tried reducing the window size and switching from CV_COMP_CORREL to CV_COMP_INTERSECT to speed up the processing, but have not seen any appreciable speed up. This may be explained by the OpenCV documentation (emphasis mine):
Each new image is measured and then
converted into an image image array
over a chosen ROI. Histograms are
taken from this image image in an area
covered by a “patch” with an anchor at
center as shown in the picture below.
The histogram is normalized using the
parameter norm_factor so that it may
be compared with hist. The calculated
histogram is compared to the model
histogram; hist uses The function
cvCompareHist() with the comparison
method=method). The resulting
output is placed at the location
corresponding to the patch anchor in
the probability image dst. This
process is repeated as the patch is
slid over the ROI. Iterative histogram
update by subtracting trailing pixels
covered by the patch and adding newly
covered pixels to the histogram can
save a lot of operations, though it is
not implemented yet.
This leaves me with a few questions:
- Is there another library that supports iterative histogram updates?
- How significant of a speed-up should I expect from using an iterative update?
- Are there any other techniques for speeding up this type of operation?
As mentioned in OpenCV Integral Histograms will definitely improve speed.
Please take a look at a sample implementation in the following link
http://smsoftdev-solutions.blogspot.com/2009/08/integral-histogram-for-fast-calculation.html