skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: On Asymmetric Classifier Training for Detector Cascades

Abstract

This paper examines the Asymmetric AdaBoost algorithm introduced by Viola and Jones for cascaded face detection. The Viola and Jones face detector uses cascaded classifiers to successively filter, or reject, non-faces. In this approach most non-faces are easily rejected by the earlier classifiers in the cascade, thus reducing the overall number of computations. This requires earlier cascade classifiers to very seldomly reject true instances of faces. To reflect this training goal, Viola and Jones introduce a weighting parameter for AdaBoost iterations and show it enforces a desirable bound. During their implementation, a modification to the proposed weighting was introduced, while enforcing the same bound. The goal of this paper is to examine their asymmetric weighting by putting AdaBoost in the form of Additive Regression as was done by Friedman, Hastie, and Tibshirani. The author believes this helps to explain the approach and adds another connection between AdaBoost and Additive Regression.

Authors:
 [1]
  1. ORNL
Publication Date:
Research Org.:
Oak Ridge National Lab. (ORNL), Oak Ridge, TN (United States)
Sponsoring Org.:
USDOE
OSTI Identifier:
963375
DOE Contract Number:
DE-AC05-00OR22725
Resource Type:
Conference
Resource Relation:
Conference: ISCV, Lake Tahoe, NV, USA, 20061105, 20061108
Country of Publication:
United States
Language:
English
Subject:
97 MATHEMATICS AND COMPUTING; 99 GENERAL AND MISCELLANEOUS//MATHEMATICS, COMPUTING, AND INFORMATION SCIENCE; ALGORITHMS; IDENTIFICATION SYSTEMS; FACE; IMPLEMENTATION; MODIFICATIONS; TRAINING

Citation Formats

Gee, Timothy Felix. On Asymmetric Classifier Training for Detector Cascades. United States: N. p., 2006. Web.
Gee, Timothy Felix. On Asymmetric Classifier Training for Detector Cascades. United States.
Gee, Timothy Felix. Sun . "On Asymmetric Classifier Training for Detector Cascades". United States. doi:.
@article{osti_963375,
title = {On Asymmetric Classifier Training for Detector Cascades},
author = {Gee, Timothy Felix},
abstractNote = {This paper examines the Asymmetric AdaBoost algorithm introduced by Viola and Jones for cascaded face detection. The Viola and Jones face detector uses cascaded classifiers to successively filter, or reject, non-faces. In this approach most non-faces are easily rejected by the earlier classifiers in the cascade, thus reducing the overall number of computations. This requires earlier cascade classifiers to very seldomly reject true instances of faces. To reflect this training goal, Viola and Jones introduce a weighting parameter for AdaBoost iterations and show it enforces a desirable bound. During their implementation, a modification to the proposed weighting was introduced, while enforcing the same bound. The goal of this paper is to examine their asymmetric weighting by putting AdaBoost in the form of Additive Regression as was done by Friedman, Hastie, and Tibshirani. The author believes this helps to explain the approach and adds another connection between AdaBoost and Additive Regression.},
doi = {},
journal = {},
number = ,
volume = ,
place = {United States},
year = {Sun Jan 01 00:00:00 EST 2006},
month = {Sun Jan 01 00:00:00 EST 2006}
}

Conference:
Other availability
Please see Document Availability for additional information on obtaining the full-text document. Library patrons may search WorldCat to identify libraries that hold this conference proceeding.

Save / Share:
  • Silicon PIN diode hybrid arrays are proposed as the ideal building blocks for a vertex detector at an asymmetric B-factory. The two-dimensional nature of the detector segmentation allows for the maximum in confusion elimination. Fine spatial resolution, on the order of 10 {mu}m per layer, is more than adequate to resolve the displaced vertices of beauty and charm decays. A high signal-to-noise ratio allows for the thinning of the detectors, reducing multiple scattering. Time tagging within the detector permits higher background levels than could otherwise be tolerated, and on-board electronics which includes zero suppression and ghost elimination, eases downstream datamore » handling and analysis. 8 refs., 3 figs.« less
  • The linear discriminant function classifier is widely used in statistical pattern recognition. The author describes how the linear classifier algorithm can be computed using systolic arrays at the word and bit levels. The bit-level array includes a new scheme for inner product computation, giving a speed and area improvement over a previous method. 14 references.
  • The presence of sulfuric acid in combustion products of Illinois No. 6 coal was examined experimentally. In these experiments the shift in mean particle size of the humidified combustion products was used to indicate the presence and form of the H/sub 2/SO/sub 4/. By varying the oxygen partial pressure in the combustion system while holding the gas temperature, coal feed rate and cooling rate constant, it was possible to change the form of H/sub 2/SO/sub 4/ found in the effluent stream. With low pO/sub 2/'s, a free sulfuric acid component was present; while at higher pO/sub 2/'s, the sulfuric acidmore » was condensed on the surface of the submicron fume. As the pO/sub 2/ increased, the amount of H/sub 2/SO/sub 4/ condensed on the surface of the particles decreased. This is attributed to an increase in vaporization rate which results in an increase in the amount of submicron fume. For a given amount of H/sub 2/SO/sub 4/ present in the system, the larger surface area presented by the submicron fume as the pO/sub 2/ increases will result in a thinner surface layer of H/sub 2/SO/sub 4/ on the individual particles.« less
  • In this paper we will discuss a project designed to provide computer recognition of the transionospheric chirps/pulses measured by the Blackbeard (BB) satellite, and expected to be measured by the upcoming FORTE satellite. The Blackbeard data has been perused by human means -- this has been satisfactory for the relatively small amount of data taken by Blackbeard. But with the advent of the FORTE system, which by some accounts might ``see`` thousands of events per day, it is important to provide a software/hardware method of accurately analyzing the data. In fact, we are providing an onboard DSP system for FORTE,more » which will test the usefulness of our Event Classifier techniques in situ. At present we are constrained to work with data from the Blackbeard satellite, and will discuss the progress made to date.« less
  • A neural network originally proposed by Szu for performing pattern recognition has been modified for use in a noisy manufacturing environment. Signals from the factory floor are frequently affine transformed and, as a consequence, a signal may not be properly aligned with respect to the input node that corresponds to the signal leading edge or with respect to the number of nodes representing the time varying part. Rater than translate and scale the presented signal, an operation which because of noise can be prone to numerical error since the signal is not smoothly varying, the network in this paper hasmore » the capability to analytically translate and scale its internal representation of the signal so that it overlays the presented signal. A response surface in the neighborhood of the stored reference signal is built during, training, and covers the range of translate and scale parameter values expected. A genetic algorithm is used to search over this hilly terrain to find the optimal values of these parameters so that the reference signal overlays the presented signal. The procedure is repeated over all hypothesized pattern classes with the best fit identifying the class.« less