OutlineIntroductionLocal FeaturesBenchmarksMikolajczyk's DatasetCaltech 101 DatasetExperiments and ResultsEvaluation of Feature DetectorsEvaluation of Feature DescriptorsFuture WorkOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkComparison of Local Feature DescriptorsSubhransu MajiDepartment of EECS,University of California, Berkeley.December 13, 2006Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture Work1IntroductionLocal Features2BenchmarksMikolajczyk’s DatasetCaltech 101 Dataset3Experiments and ResultsEvaluation of Feature DetectorsEvaluation of Feature Descriptors4Future WorkSubhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkLocal FeaturesApplications of Local FeaturesMulti Camera Scene reconstruction.Robust to Backgrounds, OcclusionsCompact Representation of Objects for Matching, Recognitionand Tracking.Lots of uses, Lots of options.This work tries to address the issue of what features aresuitable for what task, which is currently a black art!!Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkLocal FeaturesKey properties of a good local featureMust be highly distinctive, i.e. low probability of a mismatch.Should be easy to extract.Invariance, a good lo cal feature should be tolerant to.Image noiseChanges in illuminationUniform scalingRotationMinor changes in viewing directionQuestion: How to construct the local feature to achieveinvariance to the above?Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkLocal FeaturesVarious Feature DetectorsHarris detector find points at a fixed scale.Harris Laplace detector uses the scale-adapted Harris function to localizep oints in scale-space. It then selects the points for which theLaplacian-of-Gaussian attains a maximum over scale.Hessian Laplace localizes points in space at the local maxima of theHessian determinant and in scale at the local maxima of theLaplacian-of-Gaussian.Harris/Hessian Affine detector does an affine adaptation of theHarris/Hessian Laplace using the second moment matrix.Maximally Stable Exremal Regions detector finds regions such that pixelsinside the MSER have either higher (bright extremal regions) or lower(dark extremal regions) intensity than all the pixels on its outer boundary.Uniform Detector(unif) - Select 500 points uniformly on the edge mapsby rejection sampling.Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkLocal FeaturesVarious Feature DescriptorsScale Invariant Feature Transformation A local image is path is dividedinto a grid (typically 4x4) and a orientation histogram is computed foreach of these cells.Shap e Contexts computes the ditance and orientaion histogram of otherp oints relative to the interst point.Image Moments These compute the descriptors by taking various higherorder image moments.Jet Decriptors These are essentially higher order derivatives of the imageat the interest pointGradient Location and Orientaiton Histogram As the name suggests itconstructs a feature out of the image using the Histogram of location andOrientation in of points in a window around the interest point.Geometric Blur These compute the average of the edge signal responseover small tranformations. Tunable parameters include the blurgradient(β = 1), base blur (α = 0.5) and scale multiplier (s = 9).Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkLocal FeaturesExample DetectionsSubhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkMikolajczyk’s DatasetCaltech 101 DatasetEvaluation CriteriaWe want the feature to be repeatable,repeatability =correct−matchesground−truth−matchesDescriptor Performance:recall vs 1-precision graphs.recall =#correct matches#correspondancescorrect matches found by neareast neignbour matching in thefeature space.correspondances obtained from ground truth matching.1 − precision =#falsematches#false matches+#correct matcesSubhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkMikolajczyk’s DatasetCaltech 101 DatasetMikolajczyk’s Dataset8 Datasets, 6 Images per dataset.Ground Truth Homography available for these Images.Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkMikolajczyk’s DatasetCaltech 101 DatasetCaltech 101 Dataset101 Categories, man-made objects, motifs, animals and plants.Foreground Mask is available. Obtain ground truth based on arough alignement of the contours.Determine the scale, translation which maximizes area overlapof the contours.Correspondance: Features of the images within a thresholddistance(10 Pixels) under the transformation.Many clasification techniques use the structure of image forcomputing similarity. For e.g. SC based caracter recognitionusing TSP.The performance of these algorithms is dependent ondetecting features on the right positions. Ideally we wouldwant the descriptor performance to be better on such a softernotion of matching.Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkMikolajczyk’s DatasetCaltech 101 DatasetBest 8 and Worst 850 100 150 200 250 30050100150200250300100 150 20050100150200250300 350 4005010015020025050 100 150 200 2504060801001201401601802002200 50 100 150 200 25005010015020025030040 60 80 100 120 140 160 180 200 220809010011012013050 100 150 200 250 30005010015020025030040 60 80 100 120 140 160 180 200 220 2404060801001201400.9781 0.9717 0.9642 0.9490 0.9486 0.9483 0.9405 0.92230 50 100 150 200 250 300204060801000 20 40 60 80 100 120 140 160 180020406080100120140160180200100 150 200 2505010015020025050 100 150 200 250100120140160180200220100 120 140 160 180 200 220608010012014016018020022050 100 150 200 2505010015020025050 100 150 200 250 3002040608010012014016018020050 100 150 200 250 30060801001200.7097 0.6934 0.6919 0.6658 0.6614 0.6444 0.6426 0.3318Subhransu Maji Comparison of Local Feature DescriptorsOutlineIntroductionBenchmarksExperiments and ResultsFuture WorkMikolajczyk’s DatasetCaltech 101 DatasetExample Ground Truth Matches100 200 300 400
View Full Document
Unlocking...