• Documents
  • Authors
  • Tables
  • Log in
  • Sign up
  • MetaCart
  • DMCA
  • Donate

CiteSeerX logo

Advanced Search Include Citations
Advanced Search Include Citations | Disambiguate

DMCA

LOCUS: Learning Object Classes with Unsupervised Segmentation (2005)

Cached

  • Download as a PDF

Download Links

  • [johnwinn.org]
  • [www.cs.cmu.edu]
  • [www-2.cs.cmu.edu]
  • [www.cs.cmu.edu]
  • [research.microsoft.com]
  • [research.microsoft.com]
  • [www.cs.cmu.edu]
  • [www.cs.cmu.edu]
  • [www-2.cs.cmu.edu]

  • Other Repositories/Bibliography

  • DBLP
  • Save to List
  • Add to Collection
  • Correct Errors
  • Monitor Changes
by J. Winn
Venue:in ICCV
Citations:194 - 8 self
  • Summary
  • Citations
  • Active Bibliography
  • Co-citation
  • Clustered Documents
  • Version History

Citations

2120 R.: Fast approximate energy minimization via graph cuts - Boykov, Veksler, et al. (Show Context)

Citation Context

...| 2 (10) i (i,j)∈ Ē where the sums are over blocks and edges between blocks. We can find the deformation field D⋆ which maximises (10) by using graph cuts in an α-expansion algorithm, as described in =-=[18, 19]-=-. The terms in L(Q) involving the mask m take the form � 〈φ(mi)〉 Q(m) − � β ′ 〈δ(mi �= mj)〉 Q(m) (11) i (i,j)∈E where the sums are over pixels and edges between pixels. Notice that the expression now ...

1338 Active Shape Models - Their training and application. - Cootes, Cooper, et al. - 1995 (Show Context)

Citation Context

... The most labour-intensive approaches involve models whose structure is hand-crafted for a particular object class [1] or that require object-specific landmarks to be annotated on all training images =-=[2]-=-. More recent approaches use models applicable to a range of objects [3] or object classes [4, 5] but still require hand-segmented training data and so would not scale to large numbers of classes. A l...

1129 Grabcut: Interactive foreground extraction using iterated graph cuts - Rother, Kolmogorov, et al. (Show Context)

Citation Context

...enerative model incorporating both a global shape model and bottom-up edge and color cues. The part of our model used to infer the segmentation is similar to the stand-alone segmentation tool GrabCut =-=[9]-=-. However, in our case the segmentation is guided by the object class model rather than by a human user. The palette-invariance in our model builds on probabilistic index map (PIM) models [10], but LO...

1128 An introduction to variational methods for graphical models - Jordan, Ghahramani, et al. - 1999 (Show Context)

Citation Context

...the set of images I = {I1 ,...,IN }. Each image is defined by its measurement map {zi} and edge map ei. Exact inference is intractable and so we resort to using a structured variational approximation =-=[16]-=-. The approximate posterior over the variables corresponding to each image takes the factorised form P (D, T, m, λ,s|I) ≈ Q(D)Q(T)Q(λ)Q(s) � i Q(mi) and the posterior over the class and background var...

1125 Object class recognition by unsupervised scale-invariant learning - Fergus, Perona, et al. - 2003 (Show Context)

Citation Context

...ination); it does not capture variation between different objects of the same class. Now let us turn to unsupervised approaches which do not require hand-segmented training data. Constellation models =-=[7]-=- can be learned from cluttered, unsegmented images. However, due to computational restrictions, such models learn a sparse set of parts that do not cover the entire object and so do not allow for obje...

1047 What energy functions can be minimized via graph cuts - Kolmogorov, Zabih - 2004 (Show Context)

Citation Context

...| 2 (10) i (i,j)∈ Ē where the sums are over blocks and edges between blocks. We can find the deformation field D⋆ which maximises (10) by using graph cuts in an α-expansion algorithm, as described in =-=[18, 19]-=-. The terms in L(Q) involving the mask m take the form � 〈φ(mi)〉 Q(m) − � β ′ 〈δ(mi �= mj)〉 Q(m) (11) i (i,j)∈E where the sums are over pixels and edges between pixels. Notice that the expression now ...

716 Feature extraction from faces using deformable templates - Yuille, Cohen, et al. - 1993 (Show Context)

Citation Context

...arning object models can be categorised by the degree of human intervention required. The most labour-intensive approaches involve models whose structure is hand-crafted for a particular object class =-=[1]-=- or that require object-specific landmarks to be annotated on all training images [2]. More recent approaches use models applicable to a range of objects [3] or object classes [4, 5] but still require...

405 Combined object categorization and segmentation with an implicit shape model,” - Leibe, Leonardis, et al. - 2004 (Show Context)

Citation Context

...ticular object class [1] or that require object-specific landmarks to be annotated on all training images [2]. More recent approaches use models applicable to a range of objects [3] or object classes =-=[4, 5]-=- but still require hand-segmented training data and so would not scale to large numbers of classes. A less labour-intensive alternative in [6] uses motion in video to learn an object model, which is t...

191 Combined top-down/bottomup segmentation - Borenstein, Ullman - 2008 (Show Context)

Citation Context

...ticular object class [1] or that require object-specific landmarks to be annotated on all training images [2]. More recent approaches use models applicable to a range of objects [3] or object classes =-=[4, 5]-=- but still require hand-segmented training data and so would not scale to large numbers of classes. A less labour-intensive alternative in [6] uses motion in video to learn an object model, which is t...

158 Learning flexible sprites in video layers,” - Jojic, Frey - 2001 (Show Context)

Citation Context

... allows for local changes in pose and shape, the overall position and size of the object is handled by a separate global transformation T defined as a scaling S followed by a translation t. Following =-=[12]-=-, we discretise the space of transformations and restrict it to all whole-pixel translations at a small fixed range of scales. The ratio between successive scales was set at 1.2 as only a moderately c...

134 Variational message passing. - Winn, Bishop - 2005 (Show Context)

Citation Context

...rder {λ, m, T, D, π, µ o , σ o , µ b , σ b } and optimising each in turn. Rather than expand (9) for each factor separately and implement each update by hand, we use Variational Message Passing (VMP) =-=[17]-=- to apply this variational inference procedure automatically.sOriginal Inferred mask Transformed mask prob. Transformed edge mean 1 st horse in new pose Class mask probability Class edge mean Figure 2...

62 Learn to segment. In - Borenstein, Ullman - 2004 (Show Context)

Citation Context

...wever, due to computational restrictions, such models learn a sparse set of parts that do not cover the entire object and so do not allow for object segmentation. Alternatively, Borenstein and Ullman =-=[8]-=- use the overlap between automatically extracted object fragments to determine the foreground/background segmentation. However, as no global shape model is used, there is no guarantee that the resulti...

48 Extending pictorial structures for object recognition - Kumar, Torr, et al. - 2004 (Show Context)

Citation Context

...applicable to a range of objects [3] or object classes [4, 5] but still require hand-segmented training data and so would not scale to large numbers of classes. A less labour-intensive alternative in =-=[6]-=- uses motion in video to learn an object model, which is then applied to still images. However, this method only learns the variability in appearance of a single object (e.g. due to pose and illuminat...

24 Titsias. Greedy learning of multiple objects in images using robust statistics and factorial learning - Williams, K (Show Context)

Citation Context

...ble mi which is 1 if that pixel is part of the object and 0 if it is part of the background. This choice of a binary mask (rather than a real valued mask) leads to more efficient and robust inference =-=[14]-=-. The dependence of the mask on the class mask probability image π is now given by Pclass(m | π) = � i ¯πmi i (1 − ¯πi) 1−mi where ¯πi is the ith element of a deformed, transformed version of π.sWe wo...

23 Capturing image structure with probabilistic index maps, in: - Jojic, Caspi - 2004 (Show Context)

Citation Context

... GrabCut [9]. However, in our case the segmentation is guided by the object class model rather than by a human user. The palette-invariance in our model builds on probabilistic index map (PIM) models =-=[10]-=-, but LOCUS uses a more expressive color distribution in the entries of the palette, and introduces a deformation model for the index maps.s3. The LOCUS Generative Model Figure 1 shows the hierarchica...

15 Generative affine localisation and tracking - Winn, Blake - 2004 (Show Context)

Citation Context

...ersion ¯π using ¯π(x) =�π(x/S − t). (4) Although we currently restrict the global transformation to a translation and scale, we could extend our model to allow full affine transforms, as described in =-=[13]-=-. The mask model As indicated above, for each image, the ith pixel with measurement zi (a vector representing color and/or texture) has an associated mask variable mi which is 1 if that pixel is part ...

11 B.J.: Generative model for layers of appearance and deformation - Kannan, Jojic, et al. (Show Context)

Citation Context

...each object instance to vary from the canonical shape, due to within-class variability, changes in object pose or small changes in viewpoint. Our model has some similarities with that of Kannan et al =-=[11]-=- except that the deformation field is constrained to be smooth and overlapping patches are not used. We divide the image into discrete blocks (of 5 × 5 pixels) and associate a deformation vector di wi...

11 Slightly Supervised Learning of Part-Based Appearance Models - Xie, Pérez - 2004 (Show Context)

Citation Context

...nsisting of the responses of 17 texture filters. For efficiency, the mixture components are shared between the foreground and background models and only the mixing proportions differ, as suggested in =-=[15]-=-. This sharing allows for the components to be learned once (at initialization) and from then on only the proportions of pixels in each component need to be updated for each layer, with the proportion...

5 Multiple-cause vector quantization - Ross, Zemel - 2002 (Show Context)

Citation Context

...ts within the object, corresponding to, for example, an aeroplane’s tail fin, a person’s skin or a car’s windows (see Figure 4). A PIM can be seen as an extension of a Multiple Cause Vector Quantizer =-=[20]-=- with a more general form of part appearance model. For the foreground class, instead of a single distribution over local features, we use the PIM model which assigns to every pixel i an index si. The...

2 Object-specific figure-ground segmentation - Yu, Shi - 2003 (Show Context)

Citation Context

...hand-crafted for a particular object class [1] or that require object-specific landmarks to be annotated on all training images [2]. More recent approaches use models applicable to a range of objects =-=[3]-=- or object classes [4, 5] but still require hand-segmented training data and so would not scale to large numbers of classes. A less labour-intensive alternative in [6] uses motion in video to learn an...

Powered by: Apache Solr
  • About CiteSeerX
  • Submit and Index Documents
  • Privacy Policy
  • Help
  • Data
  • Source
  • Contact Us

Developed at and hosted by The College of Information Sciences and Technology

© 2007-2019 The Pennsylvania State University