This preview shows page 1-2-3-4-5-6-7-8-9-10-11-12-85-86-87-88-89-90-91-92-93-94-95-96-97-172-173-174-175-176-177-178-179-180-181-182-183 out of 183 pages.
CS294‐43: Visual Object and Activity RecognitionProf. Trevor DarrellJan 27th: Instance Recognition and Retrieval TodayToday•SIFTSIFT• Video Googlel ll• Total Recall• Photo TourismCorrespondence Fundamental to many of the core vision problems–Recognitiong– Motion tracking– Multiview geometry Local features are the keyImages from: M. Brown and D. G. Lowe. Recognising Panoramas. In Proceedings of the ) the International Conference on Computer Vision (ICCV2003Slide credit: O. Pele, S. Thrun, J. Košecká, N. KumarLocal Features: Detectors vs. DescriptorsDetectedInterest Points/RegionsDescriptors<0 12 31 0 0 23 …><5 0 0 11 37 15 …><14 21 10 0 3 22 …>Slide credit: O. Pele, S. Thrun, J. Košecká, N. KumarIdeal Interest Points/Regions Lots of themRepeatableRepeatable Representative orientation/scaleFast to extract and matchFast to extract and matchSlide credit: O. Pele, S. Thrun, J. Košecká, N. KumarKeypoint Localizationomputinggmented CTutorialensory Augcognition T• Goals: Repeatable detectionptual and SObject ReRepeatable detection Precise localization Interesting contentPercepVisual Look for two-dimensional signal changesSlide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector [Harris88]Intuition: Search for local omputingneighborhoods where the image content has two main directions (eigenvectors).gmented CTutorialensory Augcognition Tptual and SObject RePercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector [Harris88]Intuition: Search for local omputingneighborhoods where the image content has two main directions (eigenvectors).gmented CTutorialIxIyensory Augcognition T1. Image derivatives gx(D), gy(D),ptual and SObject RePercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector [Harris88]Intuition: Search for local omputingneighborhoods where the image content has two main directions (eigenvectors).gmented CTutorialIxIyensory Augcognition T1. Image derivatives gx(D), gy(D),ptual and SObject Re2. Square of Ix2Iy2IxIyPercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CoursederivativesHarris Detector [Harris88]Second moment matrix(autocorrelation matrix):omputing())()()()()(),(22DyDyxDyxDxIDIIIIIIIgIIgmented CTutorialIy1. Image derivatives2 S f IxIyI2Iy2IxIyensory Augcognition T1. Image derivatives gx(D), gy(D),2. Square of derivativesIxIyIxIyptual and SObject Re2. Square of 3. Gaussian filter g()PercepVisual derivativesfilter g(I)g(Ix2)g(Iy2)g(IxIy)Slide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector [Harris88]Second moment matrix(autocorrelation matrix):omputing())()()()()(),(22DyDyxDyxDxIDIIIIIIIg1. Image derivativesIxIygmented CTutorialIy2. Square of derivativesIx2Iy2IxIyensory Augcognition T3. Gaussian filter g(I)g(Ix2)g(Iy2)g(IxIy)ptual and SObject Re222222)]()([)]([)()(IIIIII ))],([trace()],(det[DIDIhar4. Cornerness function – both eigenvalues are strongPercepVisual g(IxIy)222222)]()([)]([)()(yxyxyxIgIgIIgIgIghar5. Non-maxima suppressionSlide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector – Responses [Harris88]omputinggmented CTutorialensory Augcognition TEffect:A very precise ptual and SObject ReEffect:A very precise corner detector.PercepVisual 12Slide credit K. Grauman, B. Leibe AAAI08 Short CourseHarris Detector – Responses [Harris88]omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selectionomputinggmented CTutorialensory Augcognition T)),(( )),((11xIfxIfmmiiii ptual and SObject ReSame operator responses if the patch contains the same image up to scale factorH t fid di th i?PercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseHow to find corresponding patch sizes?Automatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmiiSlide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmiiSlide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmiiSlide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmiiSlide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmiiSlide credit K. Grauman, B. Leibe AAAI08 Short CourseAutomatic Scale Selection• Function responses for increasing scale (scale signature) omputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual )),((1xIfmii )),((1xIfmii Slide credit K. Grauman, B. Leibe AAAI08 Short CourseLaplacian-of-Gaussian (LoG) scale detection• Laplacian also measures bandpass contrast…•which ‘scale’ has most omputing•which scale has most ‘contrast’?gmented CTutorial)()(yyxxLLensory Augcognition Tptual and SObject Re List of(x, y, s)PercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseResults: Laplacian-of-Gaussianomputinggmented CTutorialensory Augcognition Tptual and SObject RePercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseDifference-of-Gaussian (DoG)• Difference of Gaussians as approximation of theLaplacian-of-Gaussianomputingpgmented CTutorialensory Augcognition Tptual and SObject Re-=PercepVisual Slide credit K. Grauman, B. Leibe AAAI08 Short CourseDoG – Efficient Computation• Computation in Gaussian scale pyramidomputinggmented CTutorialSampling withstep =2ensory
View Full Document