Search

28 Nov 2006: Session 3: Grouping and Matching. Thursday 24 July, 09.30-12.30 (Chair: Dr A. ... Session 4: Geometry and Statistics. Thursday 24 July, 14.00-17.40 (Chair: Dr R.

19 Jul 2006: 5. 10. 108. 64. 20. 24. 68. 10. 0. 0.05. 0.1.

29 Apr 2006: average 92.0 64.1 58.3 17.0std 7.8 9.2 24.3 8.8. video sequences of the person in arbitrary motion (signif-icant translation, yaw and pitch, ... cluded and expression variant faces from a single sample per class.PAMI, 24(6), 2002.

22 Nov 2006: Cambridge UniversityEngineering Department. University of East Anglia Seminar 24. Modelling Dependencies in Sequence Classification: Augmented Statistical Models. ... Comp.EER (%). GMM A-GMM. 128 12.17 8.62256 11.24 7.88512 11.13 7.481024 10.43† 7.31.

19 Sep 2006: Cambridge University Press, 1999. [FB81] FISCHLER M., BOLLES R.: Random sample consensus:A paradigm for model-fitting with applications to image analysisand automated cartography.CACM 24, 6 (1981), 381–395.

22 Nov 2006: HMM ML – 29.4 27.3C-Aug ML CML 24.2 –. HMM MMI – 25.3 24.8C-Aug MMI CML 23.4 –. Table 2. Classification error on the TIMIT core test ... A point of particular interest is that despite poorer statesegmentation—the sufficient statistics fix the

19 Mar 2006: 24] P. S. Penev. Dimensionality reduction by sparsification in a local-features representation of human faces. ... Ross and A. Jain. Information fusion in biometrics.PatternRecognition Letters, 24(13):2115–2125, 2003.

5 Jul 2006: Cambridge UniversityEngineering Department. Trajectory Models For Speech Processing Workshop 24. Augmented Statistical Models for Speech Recognition.

21 Mar 2006: Due to the smoothness of faces, each track corre-sponds to an appearance manifold [2, 22, 24], as illustratedin Fig. ... 24] B. Moghaddam and A. Pentland. Principal manifolds and probabilis-tic subspaces for visual recognition.PAMI, 24(6), 2002.2, 3.

22 Nov 2006: ãtj = max{atj , amin} (24). whereãtj is the floored scale factor andamin is the scale floor.In this paper,amin of 0.1 was used.

21 Sep 2006: 21] for upper bodypose estimation. In [24] it is suggested to partition the parameter spaceof a 3D hand model using a multi-resolution grid. ... range. At detection rates of 0.99 the false positiverate for the centre template is 0.24, wheras it is

5 Jul 2006: Complementary System Selection (“Random”). • Variability to systems can be obtained by varying for example:– segmentation and clustering [3]– acoustic model decision tree [24]– acoustic model context (tri/quin-phone) [4]– ... Cambridge

1 Sep 2006: 10] A. M. Martinez. Recognizing imprecisely localized, partially oc-cluded and expression variant faces from a single sample per class.PAMI, 24(6), 2002.

22 Nov 2006: Devel-opment data, dev04, was made available for this task comprising2 hours of data, 24 conversations.

30 Jan 2006: FaceDB100 64.1/9.2 73.6/22.5 58.3/24.3 17.0/ 8.8FaceDB60 81.8/9.6 79.3/18.6 46.6/28.3

19 Dec 2006: Uncertainty 1 4 16 256. Clean — 33.2. SPLICENo. 24.6 20.7 17.0 12.3FE-CMLLR 16.3 15.3 12.8 13.5.

14 Sep 2006: Pattern Analysis andMachine Intelligence, 24(6):748–763, 2002. [22] S. Obdržálek and J. ... ACM Siggraph, 2004. [24] Carsten Rother, Sanjiv Kumar, Vladimir Kolmogorov,and Andrew Blake.

22 Nov 2006: ComponentsGMM A-GMM. EER(%) minDCF EER(%) minDCF. 128 12.17 0.5014 8.62 0.3714256 11.24 0.4704 7.88 0.3467512 11.13 0.4638 7.48

1 Jun 2006: IEEE Trans. Pattern Analysis and MachineIntelligence 24, 6 (2002), 748–763. [MBSL99] MALIK J., BELONGIE S., SHI J., LEUNG T.: Tex-tons, contours and regions: Cue integration in image segmenta-tion.

13 Jan 2006: 0.11 9.47 0.09fourier 5.26 0.44 12.24 0.17 9.89 0.08 8.84 0.31.

21 Nov 2006: b. . . . . (24). One candidate for estimating the decision bound-ary is the Support Vector Machine (SVM).

15 Feb 2006: For cases like these, the algorithm of [24] couldbe used to refine the class labeling. ... In: AAAI.(2005) 1508–1513. 24. Kumar, S., Hebert, M.: A hierarchical field framework for unified context-basedclassification.

29 Sep 2006: 1.24. 1.26. 1.28. 1.3. 1.32. 1.34x 10. 3. scatterer depth (m). ... f) Window length (max=175), strain and SNRe forDRS, SNRe=12.87 and V=5.24.

21 Sep 2006: 24] and for exem-plar templates by Toyama and Blake [43]. However, it is acknowledged that“one problem withexemplar sets is that they can grow exponentially with object complexity. ... We take inspiration from Jojicet al.[24] whomodeled a video

3 Apr 2006: model. In the case of gaze tracking,the standard calibration process givesn = 80 (nl = 16);with m = 24, the S3GP takes 8s to train (24s including cal-ibration) and requires

14 Sep 2006: 24, 18]. Both systems group an image’sspatial features, performing a global annealing optimizationthat propagates the certainty at distinct person-boundaries touncertain areas where those people’s outlines are ambigu-ous. ... 24] P. Tu and J.

19 Sep 2006: CACM, 24(6):381395,1981. 3. [5] D. Goldman, B. Curless, A. Hertzmann, and S.

22 Nov 2006: Now for the priors to satisfy 22. wmKm 0 (24). with the additional constraint that at least one of the meta-component valuesis greater than zero.

3 Apr 2006: Benavente. The AR face database. TechnicalReport 24, CVC, June 1998. [14] A.

22 Nov 2006: Factor analysed hidden Markov models for. speech recognition. A-V.I. Rosti , M.J.F. Gales. Cambridge University Engineering Department, Trumpington Street, Cambridge,. CB2 1PZ, UK. Abstract. Recently various techniques to improve the correlation

14 Sep 2006: 24] andfor exemplar templates by Toyama and Blake [43].However, it is acknowledged that “one problem withexemplar sets is that they can grow exponentiallywith object complexity. ... Wetake inspiration from Jojic et al. [24] who modelleda video sequence

23 Feb 2006: Gales et al.: THE CUED BROADCAST NEWS TRANSCRIPTION SYSTEM 7. developing the Cambridge 10RT broadcast news system in1998 [24]11. ... 0.16 0.18 0.2 0.22 0.24 0.2612. 14. 16. 18. 20. 22.

22 Nov 2006: The set of parameters,Θ(sm),. 1Using this form of auxiliary function yields the same update formulae asusing the extended Baum-Welch (EBW) algorithm [24], [25]. ... Wmpem =B2D. 2m B1Dm B0β. (c)m Dm. (23). where. B2 = Σ̂m (24).

14 Sep 2006: ln = 40.80RMS= 24.85. ln = 78.35RMS= 26.56. ln = 98.72RMS = 12.67. ... ln = 30.47RMS= 13.12. ln = 54.29RMS= 24.98. Figure 6:Pose Detection:This illustrates results from applying the GMM learnt from 8k marginaland 2k joint data points with 50

5 Jun 2006: 233.2.4 State-Based Speech Enhancement. 24. 3.3 Model-Based Techniques. 243.3.1 Linear Regression Adaptation. ... Ljm(τ ) = p(qjm(τ )|YT , M) (2.24)=. 1L(YT |M) Uj (τ )cjmbjm(y(τ ))βj (τ ). where. Uj (τ ) =. . a1j , if τ = 1N1i=2.

22 Nov 2006: In addition to training GI models for BN systems theuse of Gender Dependent (GD) models has been found tobe advantageous [24]. ... 0.16 0.18 0.2 0.22 0.24 0.2612. 14. 16. 18. 20. 22.

25 May 2006: Thus with no loss of accuracy Equation 9 isrewritten in the form of Equation 24.

19 Sep 2006: 3b. A simple way of measuring the silhouette coherence using theconcept of visual hull [24] is. ... pp. 245–262, 2001. [24] A. Laurentini, “The visual hull concept for silhouettebased image understanding,”IEEE Trans.

14 Jun 2006: 145]; an empirical comparisonof techniques is provided by [24]. The N-gram model captures only local constraints and ignores higher-level structure.Many more sophisticated models have been investigated.

22 Nov 2006: This sensitivity to outliers is a well known feature of the MMI criterion [24]. ... j))}. (24). Each Gaussian component is assumed to be independent of all others.

17 Feb 2006: 24]). Briefly, we estimate multivariate Gaussian components using the ExpectationMaximization (EM) algorithm [14], initialized by k-means clustering. ... appearance. IJCV, 14:5–24, 1995.34. S. Palanivel, B. S. Venkatesh, and B Yegnanarayana.

14 Mar 2006: 18] N. Vlassis and A Likas. A kurtosis-based dynamic approach to Gaussian mixture modeling.Systems, Max, and Cybernetics – Part A: Systems and Humans, 24(9):393–399, 1999.

22 Nov 2006: Km(Oi, Oj ; λ(1)) (24)where Kl(Oi, Oj ; λ) is the dot-product of the log-likelihoodratios and Kc(Oi, Oj ; λ(1)), Km(Oi, Oj ; λ(1)), etc., ... than in the ML case. This is because the MMI base-line performs much better than the ML baseline

21 Sep 2006: Pattern Recognition Letters, 24(2003):2743–2749, 2003. 8. O. Yamaguchi, K. Fukui, and K.

21 Sep 2006: The average frame rate is 24.1 frames per second(fps)). That is to say, the system can run inreal-time.

14 Sep 2006: IEEE Trans. Pattern Analysis and Machine Intell., 24(4):509–522,April 2002. 5. M. ... Journal of Computer Vision, 48(1):9–19, June 2002. 24. J. Vermaak, A.

22 May 2006: Communications of theACM, 24(6):381–395, 1981. 15. [5] A. H. Gee, R. ... H. Berman. Engineering a freehand 3Dultrasound system. Pattern Recognition Letters, 24:757–777, 2003.

30 Mar 2006: Examples includequasistatic compression imaging [26, 29], axial shear wave imaging [32] and acoustic radiationforce imaging in both quasistatic/impulsive [24] and dynamic [2] forms. ... We substitute this into Equation 24, and rearrange to produce a