### Table 1 shows the results comparing Mean-Shift and Similarity-Measure segmentation. For comparison we show our own results using affine invariant interest-points as de- scribed in [13]. In these experiments, object categorization works best with Similarity-Measure. Mean-Shift is compa- rable to affine invariant interest-points, performing better on bikes and slightly worse on cars.

"... In PAGE 3: ... Experiments and Results Experiments were carried out in two steps. First the whole approach was tested on two datasets, with Mean- Shift and Similarity-Measure segmentation and varying val- ues for regmin(see Table1 ). We used the category cars(rear) trained versus the background images from the dataset by Fergus et al.... In PAGE 4: ... Interst Point 33.3 Table1 . Relative error on dataset cars(rear) [4] and bikes [13].... ..."

### Table 1: Relative error on dataset cars(rear) [4] and bikes [13]. For two tests, we used segmentation one with regmin = 50 and one with regmin = 250. For the third test we used affine invariant interest-points. In all cases the object categorization works best with Similarity-Measure

2004

"... In PAGE 5: ... The tests were carried out on 60 new images half belonging to the learned class and half not. Table1 shows the results comparing Mean-Shift and Similarity-Measure segmentation. For com-... ..."

Cited by 10

### TABLE IV COMPARISON OF ROC-EQUAL ERROR RATES (EQ.ERR.) AND ROC-AUC (AREA UNDER CURVE) RATES ON GRAZ-01 ACHIEVED WITH THREE SPECIFIC COMBINATIONS: AFFINE INVARIANT INTEREST POINT DETECTION WITH MOMENT INVARIANTS, DOG KEYPOINT DETECTION COMBINED WITH SIFT AS DESCRIPTION METHOD AND SIMILARITY-MEASURE-SEGMENTATION (SM) DESCRIBED BY INTENSITY DISTRIBUTIONS. Dataset Moment Invariants SIFTs SM

2004

Cited by 30

### TABLE V ROC-EQUAL-ERROR RATES OF VARIOUS SPECIFIC COMBINATIONS OF REGION EXTRACTIONS AND DESCRIPTION METHODS ON THE THREE CATEGORIES OF THE GRAZ-02 DATASET. THE FIRST AND THE SECOND COLUMN ARE OBTAINED WITH THE AFFINE INVARIANT INTEREST POINT DETECTION AND MOMENT INVARIANTS OR BASIC INTENSITY MOMENTS AS LOCAL DESCRIPTOR. THE THIRD ROW WAS ACHIEVED USING DOG KEYPOINT DETECTION AND SIFTS AS DESCRIPTION METHOD USING 300 CLUSTER CENTERS WITHIN THE K-MEANS CLUSTERING. THE LAST COLUMN SHOWS THE RESULTS OF EXPERIMENTS PERFORMED USING SIMILARITY-MEASURE-SEGMENTATION AND DESCRIPTION VIA INTENSITY DISTRIBUTIONS. Dataset Moment Invariants Basic Moments SIFTs SM

2004

Cited by 30

### Table 1. Relative error on dataset cars(rear) [4] and bikes [13]. For two tests, we used segmenta- tion one with regmin =50and one with regmin = 250. For the third test we used affine invariant interest-points. In all cases the object categoriza- tion works best with Similarity-Measure

"... In PAGE 3: ... Experiments and Results Experiments were carried out in two steps. First the whole approach was tested on two datasets, with Mean- Shift and Similarity-Measure segmentation and varying val- ues for regmin(see Table1 ). We used the category cars(rear) trained versus the background images from the dataset by Fergus et al.... In PAGE 3: ... The tests were carried out on 60 new images half belong- ing to the learned class and half not. Table1 shows the results comparing Mean-Shift and Similarity-Measure segmentation. For comparison we show our own results using affine invariant interest-points as de- scribed in [13].... ..."

### TABLE V ROC-EQUAL-ERROR RATES OF VARIOUS SPECIFIC COMBINATIONS OF REGION EXTRACTIONS AND DESCRIPTION METHODS ON THE THREE CATEGORIES OF THE GRAZ-02 DATASET. THE FIRST AND THE SECOND COLUMN ARE OBTAINED WITH THE AFFINE INVARIANT INTEREST POINT DETECTION AND MOMENT INVARIANTS OR BASIC INTENSITY MOMENTS AS LOCAL DESCRIPTOR. THE THIRD ROW WAS ACHIEVED USING DOG KEYPOINT DETECTION AND SIFTS AS DESCRIPTION METHOD USING 300 CLUSTER CENTERS WITHIN THE K-MEANS CLUSTERING. THE LAST COLUMN SHOWS THE RESULTS OF EXPERIMENTS PERFORMED USING SIMILARITY-MEASURE-SEGMENTATION AND DESCRIPTION VIA INTENSITY DISTRIBUTIONS.

2004

Cited by 30

### Table 1. The values of the combined affine and blur invariants

### Table 1: This table illustrates the effect of depth error on the computed affine invariants of the fifth point for each frames. The error is computed for 0.5m distance between the camera and cubes. The entries of the table show the percent deviation of a one sigma invariant coordinate error, relative to the unit distance in the affine coordinate frame. The three error deviations, one for each coordinate, are enclosed in parentheses. The Euler angles for the rotation of the viewing direction relative to the object are indicated as a triple along the left side of the table. Each column of the table corresponds to a different affine frame.

1995

"... In PAGE 4: ... Each of the distinct frames is inserted in the model library as a separate instance of the model and has an associated invariant index. As seen from Table1 , there is a very significant difference inindexingpowerbetween the best and worst choice of affine frame. For example, for rotation (-45, 35, 0), the first row of the table, the worst of two error deviations goes down from 10.... ..."

Cited by 1

### Table 1. Some affine Hu invariants up to 4-th order mo- ments (not complete)

2003

"... In PAGE 2: ... These invariants can be easily derived using the so called complex moments, see [4]. Some of the clas- sical Hu invariants and our new Hu invariants are given in Table1 . It is very important that these features are invariant against rotations and reflections for all three types of mo- ments (DM,LM,AM).... In PAGE 2: ...1 without the normalization of the translation by the centroid. Using these normalized moments, the features Hk of Table1 can be calculated which are now invariant to homogeneous affine transformations. Table 1 shows that it can be used low order invariants, e.... In PAGE 2: ... Using these normalized moments, the features Hk of Table 1 can be calculated which are now invariant to homogeneous affine transformations. Table1 shows that it can be used low order invariants, e.g.... ..."

Cited by 2