Spatial Weighting for Bag-of-FeaturesAuthors: Marcin Marszałek, Cordelia SchmidPresented by: Brendan YoungerBetter Bags-of-FeaturesBetter Kernels- Pyramid Match Kernel, Grauman & Darrell, 2005- Mercer Kernels, Lyu, 2005Interest Point Detection- Distinctive features from keypoints, Lowe, 2004Localization- Combined segmentation & categorization, Liebe et. al., 2004Better Bags-of-FeaturesBetter Kernels- Okay, but still no spatial informationInterest Point Detection- Uses Hough transform, so restricted set of shapesLocalization- Finds “interesting” parts okay, but can’t fill in the restSpatial weightingFeatures help other features in their neighborhoodOverview of ClassificationInterest-point detectionSIFT descriptor at each interest pointFind nearest descriptors in vocabularyCreate segmentation image based on segmentations from training setWeight each feature with segmentation imageBuild histogram of features and use SVM to classifyLocal FeaturesCorner Regions: Harris-Laplace Detector (HS)“Blob”-like Regions: Laplacian Detector (LS)128-dimensional Descriptor: Lowe’s SIFT- Normalized descriptor for illumination invarianceVocabulary: K-means clustering; 1,000 features- Classification is insensitive to choice of vocabularySegmentationFind nearby features in training dataMatch location and orientation of both featuresBlur segmentation of training image based on distance between featuresAdd blurred segmentation to computed segmentationFor each feature in test image:Lather; rinse; repeatHistograms and ClassificationUsing the segmentation, weight each featurePlace features in the bucket of the nearest vocabulary featureApply the class-specific SVM to the histogram1 2 3 4 5 6 7 8 9 105143267Object LocalizationComputed a segmentation just to classify?Use the segmentation to localize and objectImprove the localization by re-running the algorithmEach time, there are fewer background features to blur the
View Full Document