Integral Channel Features - Addendum

This document is meant to serve as an addendum to [1], published at BMVC 2009. Thepurpose of this addendum is twofold: (1) to respond to feedback we’ve received since publi-cation and (2) to describe a number of changes, especially to the non-maximal suppression,that further improve performance. The performance of our updated detection increases 5%to over 91% detection rate at 1 false positive per image on the INRIA dataset, and similarlyon the Caltech Pedestrian Dataset, while overall system runtime for multiscale detectiondecreases by 1=3 to just under 1:5s per 640 480 image.We begin by rectifying an important omission to the related work. Levi and Weiss had aninnovative application of integral images to multiple image channels quite early on, demon-strating good results on face detection from few training examples [4]. This work appears tobe the earliest such use of integral images, indeed the authors even describe a precursor tointegral histograms. Many thanks to Mark Everingham for sending us this reference.