Die Präsentation wird geladen. Bitte warten

Die Präsentation wird geladen. Bitte warten

Institute for Computer Graphics and Vision – Graz University of Technology - Austria 1 Michael Donoser WS 2011/12 AKCV KU AK Computer Vision KU – WS 2011/12.

Ähnliche Präsentationen


Präsentation zum Thema: "Institute for Computer Graphics and Vision – Graz University of Technology - Austria 1 Michael Donoser WS 2011/12 AKCV KU AK Computer Vision KU – WS 2011/12."—  Präsentation transkript:

1 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 1 Michael Donoser WS 2011/12 AKCV KU AK Computer Vision KU – WS 2011/12 Michael Donoser Raum IE02042, Inffeldgasse 16/II, 8010 Graz Graz University of Technology, Austria KU-Homepage:

2 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 2 Michael Donoser WS 2011/12 AKCV KU Vortragender Michael Donoser Institut für Maschinelles Sehen und Darstellen Inffeldgasse 16/II - Raum IE (316) Sprechstunde: Montag Uhr

3 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 3 Michael Donoser WS 2011/12 AKCV KU Ziel Umsetzung einer Methode aus dem Bereich der Object Recognition Implementierung in MATLAB Umfassende experimentelle Evaluierung Präsentation der Erkenntnisse

4 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 4 Michael Donoser WS 2011/12 AKCV KU Was ist zu tun? Auswahl eines Themengebiets Bildung 2er Gruppen Bis Sonntag 20. November, 23:59 (jedes 2er Team): –Recherche zu Themengebiet (Was gibt es? Wer hat es gemacht? Was sind die wichtigsten Referenzen?) –Genaue Festlegung der Implementierungsdetails –Gibt es fertigen Code zum Vergleich? –Wie kann man die implementierte Methode evaluieren? –Schreiben eines Poposals (ein paar Seiten), welches diese Punkte beinhaltet Bis Ende Jänner: –Implementierung in Matlab –Schreiben eines Berichts mit Schwerpunkt auf die Evaluierung! –Abschlusspräsentation (Montag, 23. Jänner: 10:00-13:00)

5 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 5 Michael Donoser WS 2011/12 AKCV KU Modus Benotung 100 Punkte 50% Implementierung 30% Bericht 20% Präsentation Awards  10 Bonuspunkte: –Beste Implementierung –Beste Präsentation –Bester Bericht

6 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 6 Michael Donoser WS 2011/12 AKCV KU Termine 20. November, 23:59 Abgabe Projektproposal von jedem 2er Team per an 21. November November Gruppenweises Treffen: Besprechung der Proposals, Besprechung Unklarheiten Sonntag, :59: –Abgabe des Matlab-Codes (als zip) per an –Abgabe des Berichts (~10 Seiten) per an Latex-Style: Montag, 23. Jänner: 10:00-13:00: Präsentation

7 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 7 Michael Donoser WS 2011/12 AKCV KU 1. Robust Multi Structure Fitting Fit multiple models to noisy and outlier disturbed data Three different methods to compare (one to implement [3]) Steps: (a) Hypotheses generation, (b) multiple model fitting [1] J-linkage: Robust fitting of multiple models, R. Toldo, A. Fusiello, ECCV 2008 [2] Robust Fitting of Multiple Structures: The Statistical Learning Approach, Chin et al, ICCV 2009 [3] Common Visual Pattern Discovery via Spatially Coherent Correspondences Hairong Liu, Shuicheng Yan, CVPR 2010

8 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 8 Michael Donoser WS 2011/12 AKCV KU 2. Multi-Label Image Segmentation Segment an image in a supervised manner Dense descriptor calculation Random Forest Classification + Graph Cuts regularization Code for both algorithms available, goal: fusing [1] Interactive Foreground Extraction using Iterated Graph Cuts, Rother et al. SIGGRAPH 2004 [2] Semantic Classification by Covariance Descriptors Within a Randomized Forest, Kluckner et al., 3dRR-09 [3] Interactive Multi-Label Segmentation, Santner et al, ACCV 2010

9 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 9 Michael Donoser WS 2011/12 AKCV KU 3. Objectness Category Independent Object Localizations Provides hypotheses concerning object bounding boxes (Code avilable) Consider different saliency detectors, edges instead of segments,... [1] What is an object? Alexe, Deselaers, Ferrari, CVPR 2010 [2] Category Independent Object Proposals, Endres and Hoiem, ECCV 2010

10 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 10 Michael Donoser WS 2011/12 AKCV KU 4. Convex Graph Matching Matching a reference graph to the current frame (tracking) using a convex formulation Use SIFT points and descriptors to define the graph and apply it for tracking Code for convex matching available [ 1] H. Li, E. Kim, X. Huang, and L. He, "Object Matching with A Locally Affine-Invariant Constraint", CVPR 2010 [2] Lowe, David G, Object recognition from local scale-invariant features, ICCV 1999

11 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 11 Michael Donoser WS 2011/12 AKCV KU 5. Descriptor BRIEF Simple, efficient and effective local patch description Implementation of paper variant + Extensions: Dense Calculation for entire image Encoding Shape by Distance Transform Extension Implementation from scratch [1] BRIEF: Binary Robust Independent Elementary Features, Calonder, Lepetit, Strecha, Fua, ECCV 2010 [2] Distance Transform Templates for Object Detection and Pose Estimation, Holzer, Hinterstoisser, Ilic, Navab, CVPR 2009

12 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 12 Michael Donoser WS 2011/12 AKCV KU 6. Zero Shot Learning Learn to recognize object without any training data by defining attributes Implementing two variants of [1] Using random forest classifier [1] Learning To Detect Unseen Object Classes by Between-Class Attribute Transfer, Lamperg et al., CVPR 2009 [2] L. Breiman, Random Forests, Machine Learning, 2001

13 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 13 Michael Donoser WS 2011/12 AKCV KU 7. Patch Match Image Labelling Semantic Segmentation: Classifying local descriptors by learning from training data Analyze possibility of fast NN search to assign label patches Max-Voting assigns labels [1] The Generalized PatchMatch Correspondence Algorithm, Barnes et al., ECCV 2010 [2] Semantic Texton Forests for Image Categorization and Segmentation; Shotton et al., CVPR 2008

14 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 14 Michael Donoser WS 2011/12 AKCV KU 8. Semi-Supervised Learning Semi-Supervised Learning Standard approach: using spectral methods [1] Analyze possibility of using a game theoretical approach [2] instead [1] Semi-supervised Learning in Gigantic Image Collections, Fergus et al., NIPS 2010 [2] Common Visual Pattern Discovery via Spatially Coherent Correspondences Hairong Liu, Shuicheng Yan, CVPR 2010

15 Institute for Computer Graphics and Vision – Graz University of Technology - Austria 15 Michael Donoser WS 2011/12 AKCV KU Termine 20. November, 23:59 Abgabe Projektproposal von jedem 2er Team per an 21. November November Gruppenweises Treffen: Besprechung der Proposals, Besprechung Unklarheiten Sonntag, :59: –Abgabe des Matlab-Codes (als zip) per an –Abgabe des Berichts (~10 Seiten) per an Latex-Style: Montag, 23. Jänner: 10:00-13:00: Präsentation


Herunterladen ppt "Institute for Computer Graphics and Vision – Graz University of Technology - Austria 1 Michael Donoser WS 2011/12 AKCV KU AK Computer Vision KU – WS 2011/12."

Ähnliche Präsentationen


Google-Anzeigen