I looked into OpenCV 1.0 source code, and also 2.4.8 source code.
I saw that the program both scale image and recompute integral image.
This is contrary to Viola and Jones' ideas in their original paper.
Why is that? Is this implemented this way in OpenCV.
Here is a similar post:
http://opencv-users.1802565.n2.nabble.com/Cascade-detection-scaling-the-image-recomputing-integral-image-td7580735.html