### Table 1: Error statistics for #0Cxed-scale and multi-scale techniques for the OFCE based on the error measure quot;

1998

"... In PAGE 22: ... #5B7#5D, which is based on the spatiotemporal orientation of the measured vector v e relative to that of the correct vector v c #28recall that v = #281; u; v#29#29: quot; = arccos#28b v c #01 b v e #29 with b v = v p v #01 v : #2828#29 Although wehave maintained con#0Cdence measures with velocity estimates, wehave used a threshold on the Frobenius norm to discard uncertain vectors. In Table1 the results are listed for various spatiotem- poral scales. The following conclusions can be drawn: #0F First order approximation does not necessarily perform better than zeroth order for #0Cxed scales.... ..."

Cited by 18

### Table 3: Performance of the scale selection method when performing simultaneous spatial and scale selection based on scale-space maxima of the normalized Laplacian response using different types of hybrid multi-scale representations and either lp-normalization or variance- based normalization.

2003

Cited by 10

### Table 4: Measures of the spatial localization error when performing simultaneous spatial and scale selection based on scale-space maxima of the normalized Laplacian response using different types of hybrid multi-scale representations and either lp-normalization or variance- based normalization.

2003

Cited by 10

### Table 1. Mean squared error before filtering, after multi- scale Kalman smoothing, and after adaptive filtering. H = f0; 1g represents all pixels at the finest scale, H = f1g rep- resents pixels where LIDAR was present, and H = f0g represents pixels where LIDAR was not present.

"... In PAGE 3: ...LIDAR are summarized in Table1 for a portion of the study area. The error is the difference between the estimates and the full set of LIDAR data.... ..."

### Table 1: Parameters and results of the multi-scale algorithm.

"... In PAGE 9: ...B4CZB5 D6CTCU (in seconds); the number of candidates AC AC CB B4CZB5 AC AC after pruning and merging; and the number AC AC CB B4CZB5 CKCC AC AC of true candidates in that set. Table1 below shows the parameters and results of the test. The minimum length C4 D1CXD2 of candidates to look for was set at 210 pixels (17.... ..."

### Table 1: Parameters and results of the multi-scale algorithm.

"... In PAGE 9: ...D6CTCU (in seconds); the number of candidates AC ACCBB4CZB5AC AC after pruning and merging; and the number AC ACCBB4CZB5 CK CC AC AC of true candidates in that set. Table1 below shows the parameters and results of the test. The minimum length C4D1CXD2 of candidates to look for was set at 210 pixels (17.... ..."

### Table 5.1: Results with multi-scale feature extraction on IRMA. Multi-scale Probability model Error

### TABLE III. Comparison of classification results of the MRF and VZ MR8 classifiers for scaled data. Models are learnt either from the original textures only or the original + scaled textures while classifying both texture types. In each case, the performance of the MRF classifier is at least as good as that using the multi-scale MR8 filter bank.

Cited by 1

### Table 2. The effect of using a combination of feature types on test equal error rate. Key: KB = Kadir amp; Brady; MSH = Multi-scale Harris; C = Curves. All models had 6 parts and 40 detection/feature-type/image. Figure in bold is combination automatically chosen by train- ing/validation set.

2005

"... In PAGE 11: ... (b) Test equal error rate versus the number of detections/feature- type/image, N, for 8 part star models. In both cases the combinations of feature-types used was picked for each dataset from the results in Table2 and xed. 3.... ..."

Cited by 56

### Table 2: The effect of using a combination of feature types on test equal error rate. Key: KB = Kadir amp; Brady; MSH = Multi-scale Harris; C = Curves. All models had 6 parts and 40 detection/feature-type/image. Figure in bold is com- bination automatically chosen by training/validation set.

2005

"... In PAGE 5: ....3. Heterogeneous part experiments Here we xed all models to use 6 parts and have 40 detections/feature-type/frame. Table2 shows the different combinations of features which were tried, along with the best one picked by means of the training/validation set. We see a dramatic difference in performance between differ- ent feature types.... In PAGE 6: ... (b) Test equal error rate versus the number of detections/feature-type/image, N, for 8 part star mod- els. In both cases the combinations of feature-types used was picked for each dataset from the results in Table2 and xed. 3.... ..."

Cited by 56