1 / 35

Edges/curves /blobs

Edges/curves /blobs. Image Parsing. Regions. curve groups. Sketches coding. Rectangles recognition. Objects recognition. segmentation. segmentation. Image Parsing: Regions, curves, curve groups, curve trees, edges, texture. 2008/11/5 Hueihan Jhuang. Zhuowen Tu and Song-Chun Zhu ,

maite-haney
Télécharger la présentation

Edges/curves /blobs

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. Edges/curves /blobs

  2. Image Parsing Regions curve groups Sketches coding Rectangles recognition Objects recognition segmentation segmentation

  3. Image Parsing: Regions, curves, curve groups, curve trees, edges, texture 2008/11/5 Hueihan Jhuang

  4. Zhuowen Tu and Song-Chun Zhu , Image segmentation by Data-Driven Markov Chain Monte Carlo , PAMI,2002

  5. Objective Regions curve groups Sketches coding Rectangles recognition Objects recognition segmentation segmentation * Boundary is not necessary rectangle

  6. Segmentation • Non-probabilistic: k-means(Bishop, 1995), mean-shift(Comaniciu and Meer, 2002), agglomerative(Duda et al, 2000) • Probabilistic: GMM-EM(Moon, 1996), DDMCMC (Tu and Zhu, 2002) • Spectral factorization (Kannan et al, 2004; Meila and Shi, 2000; Ng et al, 2001; Perona and Freeman , 1998; Shi and Malik, 2000; Zass and Shashua 2005)

  7. Bayesian Formulation -notation • I: Image • K: # regions • Ri:the i-th region i = 1..K • li: the type of model underlying Ri • Θi: the parameters of the model of Ri • W: region representation Ex: K = 11, l1=U(a,b) l2=N(u,σ)

  8. Model type Number of regions Model parameters Bayesian Formulation -prior Ex: region   Region size     Region boundary θ   Θαβ

  9. I1~N(0,1) I2~N(1,2) I3~U(0,1) Bayesian Formulation -likelihood Ex: K = 11, l1=U(a,b) l2=N(u,σ) p( I | W) = N(I1;0,1) x N(I2;1,2) x …

  10. Proposed Image models Uniform Clutter Texture Shading Intensity Histogram Gabor Response Histogram (Zhu, 1998) Gaussian B-Spline Examples:    

  11. Objective in Bayesian Formulation –Maximize posterior W* = arg max p( W| I)αarg maxp( I | W) p( W ) Given a image, find the most likely partition

  12. clutter uniform texture clutter Solution space A possible solution K regions x ways to partition the image x 4 possible models for each of the k regions Enumerate all possible solutions  Monte Carlo Markov Chain 

  13. Monte Carlo Method • A class of algorithms that relies on repeat sampling to get the results • Ex: • f(x)g(x)h(x) (known but complicated probability density function) • Expectation? Integration ? maximum ? • sol: Draw a lot of samples->mean/sum/max

  14. Markov Chain Monte Carlo • MCMC is a technique for generating fair samples from a probability in high-dimensional space. Zhu 2005 ICCV tutoiral

  15. Design transitions (probability) to explore the solution space model adaptation merge Boundary diffusion split switching model The papers designs all the transition probabilities

  16. Results image Region segmentation Assume images consist of 2D compact regions, results become cluttered in the presence of 1D curve objects Proposed segmentation Zhu 2005 ICCV tutoiral

  17. Results images Segmentation using MCMC Curve groups regions

  18. Zhuowen Tu and Song-Chun Zhu, Parsing Images into Regions, Curves, and Curve Groups, IJCV, 2005.

  19. Define new components-curve family Tu and Zhu, 2005

  20. Bayesian formulation C: curve Pg: parallel curves T: curve trees p(I | Wr) -> as (Tu and Zhu, 2002), Gaussian/ historam/B-sline models p(I | Wc) -> B spline model * Minor difference from previous setting

  21. Length of curve # pixels in curve Width of the curve } exp{ Prior of curve / curve group/curve trees Attributes for each curve: Center, orientation, Length, width p(pg)~ exp{ -difference of attributes } Ex    Compatibility of curves: Width, orientation continuity,end point gap p(t)~ exp{ -incompatibility }

  22. Solution space is larger region Curve trees Free curves Curve groups MCMC again

  23. Results Less cluttered, but still doesn’t provide a good start for object recognition Tu et al, 2005

  24. Results (Tu et al, ICCV, 2003)

  25. Objective Regions curve groups Sketches coding Rectangles recognition Objects recognition segmentation segmentation

  26. Image may be divided into two components texture (Symbolic) structure (Texton, token, edges) (Julesz, 1962,1981; Marr, 1982)

  27. Cheng-en Guo, Song-Chun Zhu, and Ying Nian Wu Primal sketch: Integrating Texture and Structure , PAMI, 2005

  28. Example Edges/ blobs + + + = Likelihood? Gaussian? Histogram? + + Likelihood? Histogram as (Tu and Zhu, 2005)

  29. Proposed Image primitives

  30. Bayesian Formulation Image primitives Likelihood of non-sketchable Likelihood of sketchable priors K is the number of primitives, not # regions : this is not segmentation

  31. Sketch pursuit algorithm Sequentially adding edges that have gain more than some threshold What threshold? Blob/edge detector LOG, DG, D2G

  32. Adding one stroke into sketchable group gain

  33. Sketch pursuit algorithm For each local region, check if using other image primitives increase The posterior probability

  34. Results

  35. More Results

More Related