Search

9 Dec 2006: Rules with probability over a certain threshold are run simultaneously on the test data. ... Rules whose probability exceeds a threshold are then applied to the test data.

22 Nov 2006: The final un-adapted performance on the dev04PE test set was 41.6%. ... Table 5 shows the effect of theuse of an automatic segmenter on the dev04 test data.

22 Nov 2006: The finalsystem shows state-of-the-art performance over a range of test sets. ... This approach was not found to perform reliably across differ-ent types of test data.

21 Sep 2006: A detailed report of this test can be found in ourprevious work [10]. ... 20.2 fps).Figure 3 illustrates the recognition process on a typical test-ing clip.

9 Dec 2006: and probabilitiesFind Ngram rules. human transcriptionand diarisation. (optional)assign categories. test datatraining data. ... Rules whose probability ex-ceeds a threshold are then applied to the test data.

22 Nov 2006: Though gains inthe posterior of the correct speaker were obtained in training thesedid not generalise well to the test data. ... Eurospeech, 1997. [11] A. Martin, “The NIST year 2002 speaker recog-nition evaluation plan,” 2002, Available

21 Sep 2006: We used 18randomly selected training/test combinations for reporting identification rates. Comparative Methods. ... 0.9. 1. Dimension. Iden. tific. atio. n ra. te. Effect of the dimension on the test set.

22 Nov 2006: C-Aug ML CML 7.3 9.1. Table 1. Training and test error rates for CAN/CAN’T. ... model (tests on MMI HMMssuggest that this may yield a gain of up to 0.5% absolute).

30 Jan 2006: State-of-the-art commercial system FaceIt by Identix[12] (the best performing software in the most recentFace Recognition Vendor Test [13]),. • ... KLD) [14]. In all tests, both training data for each person in the gallery,as well as test data,

21 Sep 2006: Iden. tific. atio. n ra. te. Effect of the dimension on the test set. ... Anindependent illumination set with both training and test sets was exploited for the val-idation.

23 Feb 2006: The finalsystem shows state-of-the-art performance over a range of test sets. ... This approach was not found to perform reliably across differ-ent types of test data.

21 Sep 2006: cluttered background, and background with skin colour). The overallaccuracyon 1025 test cases is 89.7%. ... Thepercentage of test cases that cannot be mapped into any classis 20.3%.

9 Dec 2006: and probabilitiesFind Ngram rules. human transcriptionand diarisation. (optional)assign categories. test datatraining data. ... Rules whose probability ex-ceeds a threshold are then applied to the test data.

22 Nov 2006: The various models andrecognition configurations are evaluated using several recent BNdevelopment and evaluation test sets. ... Test set # Shows Hours Period. dev03 6 3 Jan. 2001eval03 6 3 Feb.

22 Nov 2006: Only 0.3% improvement wasobtained on this test set. 3See http://htk.eng.cam.ac.uk/docs/cuhtk.shtml4Significance tests were carried out using the NIST Scoring Toolkit. ... These systems wereevaluated on thedev03 andeval03 test sets, each consistingof 3

5 Jul 2006: Evaluation on held-out data (eval03)– 6 hours of test data– decoded using LVCSR trigram language model– baseline using confusion network decoding. ... φcn(O; λ) =[ F(ω1) F(ω2). φ(O; λ). ]. • Incorporating in score-space requires consistency

29 Apr 2006: To establish baseline performance, we compared ourrecognition algorithm to:. • State-of-the-art commercial system FaceItr by Identix[8] (the best performing software in the recent FaceRecognition Vendor Test [10]),. • ... perform well if imaging

15 Feb 2006: e) A test input with textons t1 and t2 in the same relative positionas that of training. ... Quantitative evaluation. Figure 8 shows the confusion matrix obtained byapplying our algorithm to the test image set.

22 Nov 2006: 3. INCREMENTAL BAYESIAN ADAPTATION. The Bayesian adaptation discussed in section 2 runs in a batchmode where all test data are assumed to be available before adap-tation. ... Theperformance was evaluated on the 2003 evaluation test dataset,eval03,

22 Nov 2006: Evaluation on held-out data (eval03)– 6 hours of test data– decoded using LVCSR trigram language model– baseline using confusion network decoding. ... φcn(O; λ) =[ F(ω1) F(ω2). φ(O; λ). ]. • Incorporating in score-space requires consistency

21 Sep 2006: The parameters for both methods are setby testing the classification performance on a test setof 5000 images. ... In a first approach,the edge and colour cost terms are computed for a numberof test images.

3 Apr 2006: In [14], error is computed using a“leave-one-out” test rather than with completely new testdata. ... A leave-one-out test for gaze-tracking data with theS3GP gives an error of 0.68.

22 Nov 2006: This is used both within a cross-validationframework and with a held-out test set, the eval03 dataset [20]. ... This is easilydemonstrated on the 6-hour eval03 test set [20]. SVMswere trained to disambiguate ten of the most confusablepairs.

17 Feb 2006: In AFR tests, such methods are usuallyoutperformed by methods from the third class: view-based techniques e.g. ... KLD) [38]. In all tests, both training data for each person in the gallery, as well as test data,consisted of only a single sequence.

22 Nov 2006: Systems aretrained using the 296 hours switchboard data (h5etrain03 )and evaluated on a 3-hour test set (dev01sub ). All systemsin this paper used 12 PLP coefficients with theC0 term plusthe first, ... 1Significance tests were carried out using the NIST

19 Mar 2006: This is challenging across large posevariations, such as those contained in our test set. ... Face recognition vendor test 2002. Technical report,National Institute of Standards and Technology, 2003.

22 Nov 2006: Thus, the final fea-ture vector has 42 dimensions. Results are given on two test sets:dev04f consists of 0.5 hours of CCTV data from shows broad-cast in November ... Thiseffect is seen for both complementary models, on both test sets.

22 Nov 2006: Table 1: Clean, matched andSPLICE on AURORA 2.0 test set A,averaged across N1-N4, WER(%). ... M-Joint1 2.43 3.82 6.97 17.1416 1.95 2.80 4.23 9.89. Table 2: Model-basedJoint systems’ performance on AURORA2.0 test set A, averaged

18 Sep 2006: create a set of such images for their tests in [7]), avideo taken of the object in its environment (e.g. ... The first setting was used as training, with the others used as test sets.

19 Dec 2006: When models trainedin clean conditions are used in the real world, the mismatchbetween the training conditions and the test causes significantloss in recognition accuracy. ... Here. P (šn|xt, M̌) P (šn|yt, M̌) (10). where the model M̌ is now

5 Jun 2006: The secondtest set was based on the 1000 word DARPA Resource Management test set. ... 71. 8.1 Average SNR for the RM test sets adding Lynx Helicopter noise attenuatedby 20dB.

3 Apr 2006: The trees are built in a simple, greedy fashion, wherenon-terminal node tests are chosen from a set of candidatefeatures together with a set of candidate thresholds to max-imise the ... and used for training wereclean, single-instance images, and so our

23 Feb 2006: The various models andrecognition configurations are evaluated using several recent BNdevelopment and evaluation test sets. ... Test set # Shows Hours Period. dev03 6 3 Jan. 2001eval03 6 3 Feb.

22 May 2006: In Section 3, we describe experiments with a hand-held probe designedto test the capabilities and limitations of the new algorithms. ... M. Rubin. Determination of scan plane motionusing speckle decorrelation: theoretical considerations and initial test.

23 Feb 2006: All test data are assumed to be available fordecoding in a single block. ... However, in some real world appli-cations test data often become available gradually.

25 May 2006: WB (t) =. π |φs2(t d̃k) φs1(t)|. π. n. (25). We test the simple strategy of Equation 25, where n determines the severity of the deweighting.WPS variants with n ... We test two families of deformation estimators. The first is an example of

30 Mar 2006: Tests of adaptive strain estimation on uniformstrain simulations are therefore expected to be independent of the AM effect. ... Fair comparison is made possible by fixing the window parameters across all of the estimatorsin each test.

14 Jun 2006: Note that the bandwidth of the test data is fixed (wide-band)across all three tests in the first column. ... Again, the bandwidth of the test data is fixed (narrow-band) across all three tests in this second column.

21 Sep 2006: we can start from an initially small model and test the ‘relevance’ ofeach new input vector i sequentially. ... The testing sets of all five pairs were used to test the system.

21 Sep 2006: This meanswe can start from an initially small model and test the ‘rele-vance’ of each new input vectori sequentially.

14 Sep 2006: density. 3 Results. We performed a number of experiments to test the method. ... The fit to the ground-truth distribution was measured as the total negativelog likelihood of 500 test data points.

21 Nov 2006: The held-out dataset for these experiments was theeval03 test set, which consists of 6 hours of data. ... To examine performance on held-out data theeval03 test set was used.

1 Sep 2006: 1-to-N and 1-to-1 matching scenarios. In the former case,we assumed that test data corresponded to one of people inthe training set and recognition was performed by

19 Sep 2006: Figure 6. Mesh evolution converges even with poor initialisation. To test the radius of convergence of the iterative mesh evolutionalgorithm the mesh was initialised from the visual hull generated by four ... The nextexperiment was designed to test this

14 Sep 2006: Ideally, human blood wouldhave been used to test our algorithm, but we experimented.

22 Jun 2006: The phantom included spheres of 1.5 cm, 1.0 cm and 0.5 cm diameter, in order to test theresolution of the algorithms. ... For this reason, it is not possible to use Field simulations created in thisway to test the frequency-based methods of attenuation

13 Jan 2006: M. Rubin. Determination of scan-plane motionusing speckle decorrelation: Theoretical considerations and initial test.

14 Sep 2006: Experimental resultsshow that the proposed technique produces good estimates of the human 3Dpose on a range of test videos in a cluttered environment.

22 Nov 2006: side. Vocal tract length normalisation (VTLN) was applied in both train-ing and test.

15 Feb 2006: the generation of standard test sequences for evaluation ofthe algorithm will be necessary.