1 / 22

Outline

Outline. Overview Integrating Vision Models CCM: Cascaded Classification Models Learning Spatial Context TAS: Things and Stuff Descriptive Querying of Images LOOPS: Localizing Object Outlines using Probabilistic Shape Future Directions. [Heitz et al. NIPS 2008b]. Image Queries on Objects.

garth
Download Presentation

Outline

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Outline • Overview • Integrating Vision Models • CCM: Cascaded Classification Models • Learning Spatial Context • TAS: Things and Stuff • Descriptive Querying of Images • LOOPS: Localizing Object Outlines using Probabilistic Shape • Future Directions [Heitz et al. NIPS 2008b]

  2. Image Queries on Objects • Categorization • Localization • Descriptive • What color is his tail? • Where is his head? • Is he… standing? sitting? bent over? • What is he doing?

  3. Related Work Boosted Detectors [ Torralba et al., PAMI 2007 ] [ Opelt, ECCV 2006 ] [ Fei-Fei and Perona, CVPR 2005 ] Coarse Refined Localization “Parts” Models [ Fergus et al., CVPR 2003 ] [ Leibe et al, ECCV 2004 ] [ Bar-Hillel et al, CVPR 2005 ] [ Winn & Shotton, CVPR 2006 ] Localization Models [ Kumar et al., CVPR 2005 ] [ Cootes et al., CVIU 1995 ] [ Borenstein et al., CVPR 2004 ] Fine OUR WORK

  4. Shape Representation: Landmarks Set of “keypoint” landmarks Shape defined by connecting piecewiselinear contour Internal landmarks are allowed (but not shown here)

  5. Training Data • Images + Outlines

  6. State-of-the-art Alternatives • kAS Detector: Edge-based object detector • Pro: No outline required Great at detection • Con: No single outline • OBJ CUT: Object-based segmentation • Pro: Produces outlines • Con: Appearance modelbased on internal texture [ Ferrari et al., CVPR 2007 ] [ Kumar et al., CVPR 2005 ] [ Prasad et al., CVPR 2006 ]

  7. LOOPS Pipeline Images + Outlines ConsistentOutlines LOOPS Model Localized Test Outlines Up Down +1 std UP -1 std +1 std DOWN -1 std Localizing Object Outlines using Probabilistic Shape Descriptive Classification register model to images shape basedclassification learn shape& landmark detectors producecorresponded training data

  8. Corresponded Outlines Images + Outlines ConsistentOutlines Localizing Object Outlines using Probabilistic Shape • Based on existing work in medical imaging • Intuition: Arc length and curvature should remain consistent [ Hill & Taylor, BMVC 1996 ] producecorresponded training data

  9. Learning Shape & Detectors ConsistentOutlines LOOPS Model +1 std -1 std +1 std -1 std Localizing Object Outlines using Probabilistic Shape learn shape& landmark detectors

  10. Multivariate Gaussian over landmark locations Shape Model Neck Legs

  11. Landmark Detectors • Build on state-of-the-art discriminative methods for detecting “parts” or “objects” Build a detector for each landmark

  12. Registration LOOPS Model Localized Test Outlines +1 std -1 std +1 std -1 std Localizing Object Outlines using Probabilistic Shape register model to images

  13. “Registering” the Model to an Image ? ? Task: Assign each landmark l L to a pixel plP L – Assignment of Landmarks to Pixels L* = argmax Score(L | I) = argmax ShapeScore(L) + ImageScore(L | I)

  14. The LOOPS MRF pairwise image score shape score landmark detectors Registering = MAP Inference over L

  15. Outlining Full LOOPS Image Detectors Only

  16. Results Rhino Giraffe Llama

  17. kAS Detector OBJ CUT LOOPS

  18. kAS Detector OBJ CUT LOOPS

  19. LOOPS Pipeline Images + Outlines ConsistentOutlines LOOPS Model Localized Test Outlines Up Down +1 std UP -1 std +1 std DOWN -1 std Localizing Object Outlines using Probabilistic Shape Descriptive Classification register model to images shape basedclassification learn shape& landmark detectors producecorresponded training data

  20. Descriptive Classification Localized Test Outlines Up Down UP DOWN Localizing Object Outlines using Probabilistic Shape Descriptive Classification shape basedclassification

  21. Descriptive Queries Goal: Classify based on shape characteristics Is the giraffe Or 1 0.8 0.6 0.4 1 2 3 4 5 6 7 8 9 10 # Training Instances (per class) “True” shape Close this gap Boosting Accuracy RANDOM

  22. Mammals

More Related