this is for holding javascript data
chengds added textit_Efficient_Detection_Detection_of__.tex
over 8 years ago
Commit id: 63307e39a766878bbd14083c50e9fd1c62f14db8
deletions | additions
diff --git a/textit_Efficient_Detection_Detection_of__.tex b/textit_Efficient_Detection_Detection_of__.tex
new file mode 100644
index 0000000..96fa3d3
--- /dev/null
+++ b/textit_Efficient_Detection_Detection_of__.tex
...
\textit{Efficient Detection}
Detection of a given part from a new image is usually performed with a sliding window approach: a coarse or fine grid of detection points is selected, and the image is tested at each point by the detector, once for every orientation angle and scale allowed for the part (we usually are not interested in all angles or scales for pedestrians). This means extracting HOG feature vectors for many configurations of position, orientation, scale, and all the approximations introduced so far make this task very efficient, especially when we use the integral image technique.
In fact, at the end of STEP 2, instead of providing the gradient histograms, we compute their integral image, so that all sums in STEP 3 can be performed in constant time for each cell, in every configuration we wish for. If the resolution of the orientation angles matches the one in the histograms binning, we expect the least amount of information loss to happen in the approximations.
