#### DMCA

## The role of saliency and error propagation in visual object recognition. (Doctoral dissertation (1995)

Venue: | Dissertation Abstracts International |

Citations: | 4 - 0 self |

### Citations

10598 | Introduction to Algorithms
- Cormen, Stein, et al.
- 2001
(Show Context)
Citation Context ... we can handle junctions, by allowing each optimal curve to make an independent set of choices. 3.3 Optimal Substructure Shortest-paths problems generally exhibit the property of optimal substructure =-=[24]-=-.2 For every pair of vertices, vl and v2 , the optimal path between this pair of vertices (denoted by P*) must contain the optimal path between any pair of vertices on F*. There exist functions that d... |

3948 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...-curve. This allowed them to run their process to convergence, without knowing the length of the optimal curve. Amini et al. [6] applied a variant of Montanari's method to implement "active contours" =-=[56]-=-. Active contours are a top-down approach for pinpointing a desired contour in an image, given an approximate location of the contour as a starting point. Assuming a starting contour, Amini et al. use... |

3922 |
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
- Fischler, Bolles
- 1981
(Show Context)
Citation Context ...particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with the image for immediate verification (e.g., =-=[77, 22, 31, 64, 5, 52]-=-). This method has become known as alignment [52], and could potentially work much faster than constrained search, because a relatively small portion of the search space is explored. The approach is j... |

1038 |
Linear algebra and its applications
- Strang
- 2005
(Show Context)
Citation Context ...n to be polynomial time in the worst case. In practice, for problems with 1 variables and m constraints, the most common algorithm, simplex, is found to usually take time proportional to Im ' (Strang =-=[84]-=-), and many highly optimized commercial implementations of simplex exist. Our problem has 6 variables and 4n constraints when n points are matched. When the number of variables in a problem is fixed a... |

893 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...increase the length of a curve faster than its curvature (e.g., Fig. 3-4), the measure does not strictly prefer larger shapes. The total squared curvature is often used as a smoothness measure (e.g., =-=[12, 56, 82, 43]-=-). In actuality, however, k is a measure of straightness, since it penalizes any deviation from straightness and only straight lines receive zero cost. For example, a Figure 3-4: Although the total sq... |

866 |
Generalizing the Hough transform to detect arbitrary shapes
- Ballard
- 1981
(Show Context)
Citation Context ...model and image features [33], though it can be done in high-polynomial time [17]. Transformation clustering searches the space of all possible transformations between model and image features (e.g., =-=[8, 89, 88, 63]-=-). Although most implementations take low polynomial time, they are sensitive to false positives, that is, they are likely to incorrectly identify the model [34]. There are implementations of this tec... |

527 |
Perceptual Organization and Visual Recognition
- Lowe
- 1985
(Show Context)
Citation Context ...particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with the image for immediate verification (e.g., =-=[77, 22, 31, 64, 5, 52]-=-). This method has become known as alignment [52], and could potentially work much faster than constrained search, because a relatively small portion of the search space is explored. The approach is j... |

467 |
Recognition by linear combinations of models
- Ullman, Basri
- 1991
(Show Context)
Citation Context ...ate an initial match of more than three points (such as Lowe's [64], Roberts' [77], Jacobs' [55], and Wayner's [92]), and some alignment approaches create an initial alignment using only three points =-=[77, 31, 64, 52, 91]-=-. In the latter case, a recognition system might attempt to add matches, and use these additional matches to narrow the area in which it must search for still more consistent matches. Additionally, th... |

406 |
Geometric Hashing: A General and Efficient Model-Based Recognition Scheme.
- Lamdan, Wolfson
- 1988
(Show Context)
Citation Context ...hes to recognition. Beveridge et al. [10] use a robust method to evaluate particular model poses. Error propagation has also been studied in the context of Geometric Hashing approaches to recognition =-=[59, 60, 61]-=-. Costa et al. [25] considered the distribution of uncertainty regions in terms of the affine invariant parameters that describe the image points. Rigoutsos and Hummel [75, 76] also considered this is... |

357 |
Using dynamic programming for solving variational problems in Vision,”
- Amini, Weymouth, et al.
- 1990
(Show Context)
Citation Context ...o decrease according to a geometric series when evaluated around any closed-curve. This allowed them to run their process to convergence, without knowing the length of the optimal curve. Amini et al. =-=[6]-=- applied a variant of Montanari's method to implement "active contours" [56]. Active contours are a top-down approach for pinpointing a desired contour in an image, given an approximate location of th... |

333 |
Machine Perception of Three-Dimensional Solids. Outstanding Dissertations in the Computer Sciences
- Roberts
- 1963
(Show Context)
Citation Context ...particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with the image for immediate verification (e.g., =-=[77, 22, 31, 64, 5, 52]-=-). This method has become known as alignment [52], and could potentially work much faster than constrained search, because a relatively small portion of the search space is explored. The approach is j... |

278 |
Recognizing solid objects by alignment with an image
- Huttenlocher, Ullman
- 1990
(Show Context)
Citation Context |

257 |
Complex Analysis,
- Ahlfors
- 1953
(Show Context)
Citation Context ... curve is smoother. In the rightmost picture in Fig. 3-12, we tried an experiment where we incorporated the gradient magnitude at the actual elements into the cost function. Specifically, we let cj E =-=[0, 1]-=- be the normalized directional derivative across element vj. Then for every actual edge (vi, vj) in the input graph, we changed its weight from w(vi, vj) = kij to w(vi, vj) = kij + A'(1 - cj), with A'... |

257 |
Human image understanding: recent research and a theory
- Biederman
- 1985
(Show Context)
Citation Context ...r three reasons. First, the features used frequently are local, which means they are insensitive to occlusion. Non-local, region-based features such as moments [27, 74], skeletons [85, 86], and geons =-=[11]-=- are difficult to obtain if regions can be partially occluded. Even though model-based representations may include such global features, in this work we are interested in features that can be recovere... |

254 | Symbolic reasoning among 3D models and 2D images - Brooks - 1981 |

240 | Trace inference, curvature consistency and curve detection
- Parent, Zucker
- 1989
(Show Context)
Citation Context ...ponential space, but then may reduce the search in practice by using heuristics which can lead to sub-optimal solutions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal =-=[73, 53, 83, 72]-=-, or else incorporate substantial domain-specific information to constrain the possible groupings [67, 98, 68]. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that... |

209 | Inferring global perceptual contours from local features
- Guy, Medioni
- 1996
(Show Context)
Citation Context ...isions of the network's elements for their optimal neighbors. As a result, the grouping system could make choices at junctions that were not optimal according to the saliency measure. Guy and Medioni =-=[43]-=- also produced a saliency map from an edge image. In their scheme, each point in the image receives a saliency value equal to a weighted sum of contributions from the individual edge elements. The wei... |

209 | Stochastic completion fields: a neural model of illusory contour shape and saliency
- Williams, Jacobs
- 1997
(Show Context)
Citation Context ...s of our saliency approach-we are looking to provide a means to obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, =-=[44, 99, 42, 30]-=- identify occluded and subjective contours, and [90, 79, 15, 16, 48, 93, 19] determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as su... |

200 | Faster shortest-path algorithms for planar graphs.
- Henzinger, Klein, et al.
- 1997
(Show Context)
Citation Context ... running time would be O(p(V log V + E)) = O(p2b(logp + b)). Moreover, it is possible to do even better because our graph is locally connected, using the recent, linear-time algorithm of Klein et al. =-=[57]-=-. As a consequence the time for our method is bounded asymptotically by O(pE) = O(p2b2), which is the same as the time for Shashua and Ullman's method. It is worth noting that for most images, we obta... |

196 |
Linear programming and convex hull made easy. In:
- Seidel
- 1990
(Show Context)
Citation Context ...when n points are matched. When the number of variables in a problem is fixed and only the number of constraints grows, as in our case, there are algorithms that take linear expected time (see Seidel =-=[81]-=-, for example). When the errors in the image points are Gaussian distributed and there are more than three matched points, the Kalman filter can be used to recursively compute Gaussian distributions f... |

189 |
Localizing overlapping parts by searching the interpretation tree.
- Grimson, Lozano-Perez
- 1987
(Show Context)
Citation Context ...to error in the data features. One approach to recognition, called constrained search, performs a depthfirst search of this tree, backtracking whenever an inconsistent set of pairings is found (e.g., =-=[13, 32, 71, 39, 40, 46]-=-). Another recognition approach, which is particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with th... |

180 | Model-Based Recognition and Localization from Sparse Range or Tactile Data,” Inter.
- Grimson, Lozano-Perez
- 1984
(Show Context)
Citation Context ...to error in the data features. One approach to recognition, called constrained search, performs a depthfirst search of this tree, backtracking whenever an inconsistent set of pairings is found (e.g., =-=[13, 32, 71, 39, 40, 46]-=-). Another recognition approach, which is particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with th... |

178 |
Extracting straight lines.
- Burns, Hanson, et al.
- 1986
(Show Context)
Citation Context ...each curve point we would apply a dedicated line-fitting process that looks back in the intensity image at the area around the curve to find long, straight line segments (see for example Burns et al. =-=[20]-=-). The endpoints of these segments become our basis points. We suppose the output from this step looks like that shown in Fig. 6-3-center. 2. Alignment stage. The alignment stage expects as input seve... |

159 | Structural saliency: The detection of globally salient structures using a locally connected network
- Sha’Ashua, Ullman
- 1988
(Show Context)
Citation Context ...e curve over another as more salient, and then the two curves are uniformly scaled in size, the mechanism's choice should not change. Our notion of global saliency is borrowed from Shashua and Ullman =-=[82]-=-, who first demonstrated how saliency could be computed for structures like the circle and the blob in Fig. 1-3. They defined a network of locally connected elements on top of an image. Using dynamic ... |

152 | A New Approach for the Recognition and Positioning of Two-Dimensional Objects”, - Ayache, Faugeras - 1986 |

146 |
Integrating region growing and edge detection,
- Pavlidis, Liow
- 1990
(Show Context)
Citation Context ...ponential space, but then may reduce the search in practice by using heuristics which can lead to sub-optimal solutions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal =-=[73, 53, 83, 72]-=-, or else incorporate substantial domain-specific information to constrain the possible groupings [67, 98, 68]. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that... |

141 | Recognizing and locating partially visible objects: the local-feature-focus method
- Bolles, Cain
- 1982
(Show Context)
Citation Context ...to error in the data features. One approach to recognition, called constrained search, performs a depthfirst search of this tree, backtracking whenever an inconsistent set of pairings is found (e.g., =-=[13, 32, 71, 39, 40, 46]-=-). Another recognition approach, which is particularly pertinent to this thesis, only goes down to a short depth in the tree and then uses the current set of correspondences to align the model with th... |

135 |
Affine invariant model-based object recognition,
- Lamdan, Schwartz, et al.
- 1990
(Show Context)
Citation Context ...locher and Ullman [52]). Other approaches use indexing to match more than the minimal number before looking for confirming features (e.g., Rothwell et al. [78], Thompson and Mundy [88], Lamdan et al. =-=[60]-=-, Jacobs [55]). Most :recognition systems take an ad-hoc approach to the problem of accounting for the effects of sensing error on the projected positions of unmatched model features. Some systems mat... |

131 |
Object Recognition by Affine Invariant Matching.
- Lamdan, Schwartz, et al.
- 1988
(Show Context)
Citation Context ...hes to recognition. Beveridge et al. [10] use a robust method to evaluate particular model poses. Error propagation has also been studied in the context of Geometric Hashing approaches to recognition =-=[59, 60, 61]-=-. Costa et al. [25] considered the distribution of uncertainty regions in terms of the affine invariant parameters that describe the image points. Rigoutsos and Hummel [75, 76] also considered this is... |

128 | On the sensitivity of the hough transform for object recognition.
- Grimson, Huttenlocher
- 1990
(Show Context)
Citation Context ...l and image features (e.g., [8, 89, 88, 63]). Although most implementations take low polynomial time, they are sensitive to false positives, that is, they are likely to incorrectly identify the model =-=[34]-=-. There are implementations of this technique that are robust to false positives, but they require high-polynomial time [21]. The third technique is 130 alignment, which is low polynomial, and one of ... |

114 |
Three-dimensional model matching from an unconstrained viewpoint
- Thompson, Mundy
- 1987
(Show Context)
Citation Context ... grouping is "local grouping," which collects edge features that are nearby in the image. Examples of local groups include points from the same curve segment [51] and straight edges that form corners =-=[88]-=- or three-line junctions [46]. A difficulty with such methods is that local groups tend to arise accidentally, and so false-positive groups of this kind can be ubiquitous in an image. More global grou... |

97 |
Filling-in the gaps: The shape of subjective contours and a model for their creation,”
- Ullman
- 1976
(Show Context)
Citation Context ...obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, [44, 99, 42, 30] identify occluded and subjective contours, and =-=[90, 79, 15, 16, 48, 93, 19]-=- determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as such into the next four chapters. Saliency is discussed in the second and thir... |

87 | The curve of least energy,”
- Horn
- 1983
(Show Context)
Citation Context ...obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, [44, 99, 42, 30] identify occluded and subjective contours, and =-=[90, 79, 15, 16, 48, 93, 19]-=- determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as such into the next four chapters. Saliency is discussed in the second and thir... |

80 | On the Verification of Hypothesized Matches in Model-Based Recognition
- Grimson, Huttenlocher
- 1991
(Show Context)
Citation Context ...ing that the features these events introduce effectively occur at random. This assumption has been made before for analyzing the verification stage of 2D recognition, and has yielded accurate results =-=[36]-=-. It is important to understand that the event I is dependent on the recognition algorithm. For instance, for Huttenlocher and Ullman's alignment algorithm, I is the event that a particular number of ... |

72 |
Model-Based Image Matching Using Location
- Baird
- 1985
(Show Context)
Citation Context ...t aimed at better understanding the effects of error on the matching process. Some of this work attempts to design algorithms that are guaranteed to perform well in the presence of error (e.g., Baird =-=[7]-=-, 1This section is joint work between the author and David Jacobs. Cass [21], Breuel [17]), but most relevant to this thesis is work that also examines the propagation of error in recognition systems.... |

68 |
Pose Determination of a Three-dimensional Object using Triangle Pairs,”
- Linnaimaa, Harwood, et al.
- 1988
(Show Context)
Citation Context ...model and image features [33], though it can be done in high-polynomial time [17]. Transformation clustering searches the space of all possible transformations between model and image features (e.g., =-=[8, 89, 88, 63]-=-). Although most implementations take low polynomial time, they are sensitive to false positives, that is, they are likely to incorrectly identify the model [34]. There are implementations of this tec... |

68 |
On the optimal detection of curves in noisy pictures
- Montanari
- 1971
(Show Context)
Citation Context ...that is, there is no attempt to optimize a measure of saliency over the set of image curves. The use of dynamic programming to find optimal curves in images was first introduced in 1.971 by Montanari =-=[69]-=-. Montanari's method needs to know the length of the curve for which it is searching, since otherwise the best curve could be infinitely long. Shashua and Ullman's dynamic-programming method is simila... |

67 |
Figure-ground discrimination: a combinatorial optimization approach
- Herault, Horaud
- 1993
(Show Context)
Citation Context ...ings. To find desirable groups, some approaches potentially search the entire exponential space, but then may reduce the search in practice by using heuristics which can lead to sub-optimal solutions =-=[47, 26, 49, 55]-=-. Other methods use greedy algorithms which need not be optimal [73, 53, 83, 72], or else incorporate substantial domain-specific information to constrain the possible groupings [67, 98, 68]. On the p... |

60 |
Neural dynamics of surface perception: boundary webs, illuminants, and shape from shading
- Grossberg, Mingolla
- 1987
(Show Context)
Citation Context ...s of our saliency approach-we are looking to provide a means to obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, =-=[44, 99, 42, 30]-=- identify occluded and subjective contours, and [90, 79, 15, 16, 48, 93, 19] determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as su... |

57 |
On the sensitivity of geometric hashing.
- GRIMSON, HUTTENLOCHER
- 1990
(Show Context)
Citation Context ...esults for recognition schemes. Lamdan and Wolfson [62] considered the related problem of determining when three image points provide an unstable basis for Geometric Hashing. Grimson and Huttenlocher =-=[35]-=- considered the effects of bounded error on Geometric Hashing, and provided loose bounds on this effect. Jacobs [54] determined exactly how bounded error effects Geometric Hashing indices. Grimson et ... |

56 |
Special purpose automatic programming for 3-D model-based vision
- Goad
- 1983
(Show Context)
Citation Context ...correspond. Approaches to model-based recognition can be divided into three main classes. Constrained search searches the space of all possible correspondences between model and image features (e.g., =-=[13, 32, 71, 39, 40, 46]-=-). It has been shown that most implementations take exponential time in the numbers of model and image features [33], though it can be done in high-polynomial time [17]. Transformation clustering sear... |

50 | A Study of Affine Matching with Bounded Sensor Error
- Grimson, Huttenlocher, et al.
- 1992
(Show Context)
Citation Context ...e the hypothesis. Given the number of non-empty search regions, the hypothesis is accepted if and only if the number is greater than a pre-chosen threshold. The alignment algorithms examined in [36], =-=[38]-=-, and [4] evaluate a hypothesis using a threshold on the probability of a false positive, that is, of a random feature arising in each of the 133 search regions. In these algorithms, I is the event th... |

50 | Finding geometric and relational structures in an image - Horaud, Veillon - 1990 |

48 |
An Application of Heuristic Search Methods to Edge and Contour Detection
- Martelli
- 1976
(Show Context)
Citation Context ...r. In addition to dynamic programming, shortest-paths algorithms also have been used before to find curves in images. To our knowledge there are only two such works. The first was by Martelli in 1976 =-=[66]-=-. Martelli's formulation involves searching for best paths through a sequence of tables whose sizes are on the order of the size of the image. The number of tables is potentially the size of the image... |

46 |
Object recognition using three-dimensional information.
- Oshima, Shirai
- 1983
(Show Context)
Citation Context |

43 |
New methods for matching 3d objects with single perspective views,”
- Horaud
- 1987
(Show Context)
Citation Context |

41 |
A baysian multiple-hypothesis approach to edge grouping and contour segmentation.
- Cox, Rehg, et al.
- 1993
(Show Context)
Citation Context ...with more global approaches is that there are an exponential number of ways to form groups from the image edges. One approach is to perform a backtracking, constrained search through all of the edges =-=[47, 26, 55]-=-. The danger, however, is that either the search time will be prohibitive or else the constraints will be so strong that the discovered curves may not be near optimal. On the other hand, there may be ... |

41 |
Recognizing partially occluded parts
- Tumey, Mudge, et al.
(Show Context)
Citation Context ...model and image features [33], though it can be done in high-polynomial time [17]. Transformation clustering searches the space of all possible transformations between model and image features (e.g., =-=[8, 89, 88, 63]-=-). Although most implementations take low polynomial time, they are sensitive to false positives, that is, they are likely to incorrectly identify the model [34]. There are implementations of this tec... |

40 |
Three-dimensional recognition of solid objects from a two-dimensional image
- Huttenlocher
- 1988
(Show Context)
Citation Context ...to their projections. The simplest form of grouping is "local grouping," which collects edge features that are nearby in the image. Examples of local groups include points from the same curve segment =-=[51]-=- and straight edges that form corners [88] or three-line junctions [46]. A difficulty with such methods is that local groups tend to arise accidentally, and so false-positive groups of this kind can b... |

40 |
Statistical Object Recognition
- Wells
- 1993
(Show Context)
Citation Context ... that is used to match four model and image points. This is useful because, currently, the least-squares solution itself can be found only through iterative methods. For both 3D and 2D objects, Wells =-=[96, 97]-=- used a Bayesian approach and Gaussian error assumptions to derive an evaluation function that measures the likelihood of any given pose. Wells then used heuristic search and gradient descent methods ... |

39 |
Finding Convex Edge Groupings in an Image
- Huttenlocher, Wayner
- 1990
(Show Context)
Citation Context ...ponential space, but then may reduce the search in practice by using heuristics which can lead to sub-optimal solutions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal =-=[73, 53, 83, 72]-=-, or else incorporate substantial domain-specific information to constrain the possible groupings [67, 98, 68]. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that... |

37 |
On the error analysis of geometric hashing
- Lamdan, Wolfson
(Show Context)
Citation Context ...76] also considered this issue for Gaussian and uniform error. Both Costa et al. and Rigoutsos and Hummel then considered the implications of these results for recognition schemes. Lamdan and Wolfson =-=[62]-=- considered the related problem of determining when three image points provide an unstable basis for Geometric Hashing. Grimson and Huttenlocher [35] considered the effects of bounded error on Geometr... |

33 | Efficient model library access by projectively invariant indexing functions
- Rothwell, Zisserman, et al.
- 1992
(Show Context)
Citation Context ... Ayache and Faugeras [5], Horaud [46], Huttenlocher and Ullman [52]). Other approaches use indexing to match more than the minimal number before looking for confirming features (e.g., Rothwell et al. =-=[78]-=-, Thompson and Mundy [88], Lamdan et al. [60], Jacobs [55]). Most :recognition systems take an ad-hoc approach to the problem of accounting for the effects of sensing error on the projected positions ... |

30 |
The Probabilistic Peaking Effect of Viewed Angles and Distances with Application to 3-D Object Recognition
- Ben-arie
- 1990
(Show Context)
Citation Context ...ing. It will be useful to see how important is Pr[HIV] at distinguishing correct hypotheses. Its importance will depend on the effectiveness of the "probabilistic peaking effect," studied by Ben-arie =-=[9]-=- and applied to alignment by Olson [70]. The point of the probabilistic peaking effect is that, although all configurations of image triples can be produced by the model triple, some of these configur... |

30 |
Combinatorial Optimization Applied to Variable Scale 2D Model Matching
- Beveridge, Weiss, et al.
- 1990
(Show Context)
Citation Context ...ainty in every predicted point (e.g., [87]). Sarachik and Grimson [80] used this observation to propose a new method of performing and evaluating alignment approaches to recognition. Beveridge et al. =-=[10]-=- use a robust method to evaluate particular model poses. Error propagation has also been studied in the context of Geometric Hashing approaches to recognition [59, 60, 61]. Costa et al. [25] considere... |

29 |
Object discrimination based on depth-from-occlusion. Neural Comput.
- LH, Sajda
- 1992
(Show Context)
Citation Context ...s of our saliency approach-we are looking to provide a means to obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, =-=[44, 99, 42, 30]-=- identify occluded and subjective contours, and [90, 79, 15, 16, 48, 93, 19] determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as su... |

28 |
Image Chunking: Defining Spatial Building Blocks for Scene Analysis
- Mahoney
- 1987
(Show Context)
Citation Context ...salient a noisy background. Middle: curve (S = 0.5445). Figure 3-9: Smoothness and size can be important for saliency. On the left, the input image contains a globally salient blob (the image is from =-=[65]-=-). The middle and right images show the shortest-paths saliency map and best closed curve (for which S = 0.1272), respectively. be bridged to find the salient curve, and so the relevant term in the sa... |

26 |
Application of tensor theory to object recognition and orientation determination
- Cyganski, Orr
- 1985
(Show Context)
Citation Context ...s popular and has become so primarily for three reasons. First, the features used frequently are local, which means they are insensitive to occlusion. Non-local, region-based features such as moments =-=[27, 74]-=-, skeletons [85, 86], and geons [11] are difficult to obtain if regions can be partially occluded. Even though model-based representations may include such global features, in this work we are interes... |

26 | The combinatorics of heuristic search termination for object recognition in cluttered environments
- Grimson
- 1991
(Show Context)
Citation Context ... hypotheses to try. Not surprisingly, many authors have argued that for recognition of this form to be feasible, grouping is necessary as a precursor to the matching process (e.g., Lowe [64], Grimson =-=[33]-=-, Jacobs [55]). The purpose of grouping is to identify subsets of image features that are likely to come from a single object, without reference to a model database. The idea is that images of modeled... |

25 | Recognizing 3D Objects from 2D Images: An Error Analysis
- Grimson, Huttenlocher, et al.
- 1992
(Show Context)
Citation Context ...eters. Furthermore, we will derive direct expressions for the predicted points in terms of the matched points, so that we do not explicitly go through a rigid transformation. Recently, Grimson et al. =-=[37]-=- presented a formal analysis of error propagation starting from the matched image points, for three-dimensional objects. They considered scaled-orthographic projection and bounded, circular error. Sta... |

24 |
On minimal energy trajectories,”
- Bruckstein, Netravali
- 1990
(Show Context)
Citation Context ...obtain the higher-level primitives in the first place. Other work on grouping addresses different aspects of the problem. For instance, [44, 99, 42, 30] identify occluded and subjective contours, and =-=[90, 79, 15, 16, 48, 93, 19]-=- determine the shapes of gap completions. 1.4 Overview of the Thesis This thesis contains four pieces of work, divided as such into the next four chapters. Saliency is discussed in the second and thir... |

24 |
Robust Estimation of Camera Location and Orientation from Noisy Data having Outliers
- Kumar, Hanson
- 1989
(Show Context)
Citation Context ...model features have been matched to over-determine the pose. Some of this work analyzes the effect that errors in image features have on the accuracy of the resulting pose, including Kumar and Hanson =-=[58]-=- and Hel-Or and Werman [45]. The work of Hel-Or and Werman is particularly relevant to us, because they also consider how error propagates through the pose to the projections of unmatched feature poin... |

23 | Distance metric between 3-D models and 2D images for recognition and classification
- Basri, Weinshall
- 1996
(Show Context)
Citation Context ... the model is planar. Alter and Grimson's technique sometimes performed poorly when the two regions overlapped, because it had difficulty distinguishing them. Also for 3D objects, Weinshall and Basri =-=[94]-=- provided analytic bounds on the amount of error in a least-squares solution that is used to match four model and image points. This is useful because, currently, the least-squares solution itself can... |

21 |
3-D Pose from 3 Points Using Weak-Perspective
- Alter
- 1994
(Show Context)
Citation Context ...ponding points are minimal to determine the transformation. Given three corresponding points, the transformation always exists if the model points are not collinear and it generally has two solutions =-=[52, 2]-=-; in particular, the scale factor and translation are always unique, and the'rigid rotation matrix is unique up to a reflection of the rotated model about a plane parallel to the image. For 3D linear ... |

21 |
Recognizing 3-D Objects Using 2-D Images
- Jacobs
- 1993
(Show Context)
Citation Context ...st alignment and indexing approaches to recognition. Some alignment approaches use grouping methods to generate an initial match of more than three points (such as Lowe's [64], Roberts' [77], Jacobs' =-=[55]-=-, and Wayner's [92]), and some alignment approaches create an initial alignment using only three points [77, 31, 64, 52, 91]. In the latter case, a recognition system might attempt to add matches, and... |

20 |
Efficiently Using Invariant Theory for Model-based Matching
- Wayner
- 1991
(Show Context)
Citation Context ...dexing approaches to recognition. Some alignment approaches use grouping methods to generate an initial match of more than three points (such as Lowe's [64], Roberts' [77], Jacobs' [55], and Wayner's =-=[92]-=-), and some alignment approaches create an initial alignment using only three points [77, 31, 64, 52, 91]. In the latter case, a recognition system might attempt to add matches, and use these addition... |

17 | 3d shape representation by contours,”
- Weiss
- 1988
(Show Context)
Citation Context ...n this pair of vertices (denoted by P*) must contain the optimal path between any pair of vertices on F*. There exist functions that do not obey this restriction. For example, the measure proposed in =-=[95, 19]-=-, T(F) = lk, where 1 denotes the length of F and k denotes its total squared curvature, does not have optimal substructure. To see this, consider the two curves on Fig. 3-3. The curve FI is composed o... |

16 |
Optimal Affine-Invariant Point Matching
- Costa, Haralick, et al.
- 1990
(Show Context)
Citation Context ...ge et al. [10] use a robust method to evaluate particular model poses. Error propagation has also been studied in the context of Geometric Hashing approaches to recognition [59, 60, 61]. Costa et al. =-=[25]-=- considered the distribution of uncertainty regions in terms of the affine invariant parameters that describe the image points. Rigoutsos and Hummel [75, 76] also considered this issue for Gaussian an... |

16 |
Robust similarity invariant matching in the presence of noise
- Rigoutsos, Hummel
- 1991
(Show Context)
Citation Context ...ches to recognition [59, 60, 61]. Costa et al. [25] considered the distribution of uncertainty regions in terms of the affine invariant parameters that describe the image points. Rigoutsos and Hummel =-=[75, 76]-=- also considered this issue for Gaussian and uniform error. Both Costa et al. and Rigoutsos and Hummel then considered the implications of these results for recognition schemes. Lamdan and Wolfson [62... |

15 |
Fast and Robust 3D Recognition by Alignment
- Alter, Grimson
- 1993
(Show Context)
Citation Context ...f point features. The numerical technique is less practical, however, for use at run-time in a recognition system. Using the same projection and error models as Grimson et al. [37], Alter and Grimson =-=[4]-=- presented experiments that show that the true uncertainty regions tend to be circular to a good approximation, and presented a numeric method for more accurately bounding the uncertainty regions. Thi... |

15 | Massively Parallel Bayesian Object Recognition
- Rigoutsos
- 1992
(Show Context)
Citation Context ...ches to recognition [59, 60, 61]. Costa et al. [25] considered the distribution of uncertainty regions in terms of the affine invariant parameters that describe the image points. Rigoutsos and Hummel =-=[75, 76]-=- also considered this issue for Gaussian and uniform error. Both Costa et al. and Rigoutsos and Hummel then considered the implications of these results for recognition schemes. Lamdan and Wolfson [62... |

14 |
Optimal Matching of Planar Models in 3D Scenes
- Jacobs
- 1991
(Show Context)
Citation Context ...nd y coordinates of the predicted model points. The x and y error bounds are guaranteed to hold only if the matched model points are far enough apart (approximately five pixels). Subsequently, Jacobs =-=[54]-=- showed that the true uncertainty regions are discs, and gave analytic expressions for their centers and radii. These regions are circular because in this case the projection model is linear in such a... |

14 |
Gaussian Error Models for Object Recognition
- Sarachik, Grimson
- 1993
(Show Context)
Citation Context ...m Gaussian. But if the error is believed to be Gaussian, then each candidate image feature could be ranked according to how likely the model feature was of producing it. In fact. Sarachik and Grimson =-=[80]-=- gave a method which aggregates such rankings into a single measure. Using the measure, their method decides whether to accept or reject the hypothesis based on a pre-determined, formally chosen thres... |

13 |
Space and time bounds on model indexing
- Clemens, Jacobs
- 1991
(Show Context)
Citation Context ...h it must search for still more consistent matches. Additionally, the algorithm from Section 4.7 may be useful in methods that match image to model features by indexing, and then verify these matches =-=[60, 23, 55, 88, 78, 92]-=-. In these approaches, some model features are matched to image features to determine a model pose, and then this pose is used to find matches for additional model features. Our results show exactly w... |

12 |
MAP Model Matching
- Wells
- 1991
(Show Context)
Citation Context ... that is used to match four model and image points. This is useful because, currently, the least-squares solution itself can be found only through iterative methods. For both 3D and 2D objects, Wells =-=[96, 97]-=- used a Bayesian approach and Gaussian error assumptions to derive an evaluation function that measures the likelihood of any given pose. Wells then used heuristic search and gradient descent methods ... |

12 |
Perceptual Organization of Occluding Contours
- Williams
- 1990
(Show Context)
Citation Context ...tions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal [73, 53, 83, 72], or else incorporate substantial domain-specific information to constrain the possible groupings =-=[67, 98, 68]-=-. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that encourages long and straight curves. Using dynamic programming, they implemented a relaxation process on a ne... |

11 |
The Perception of Subjective Surfaces
- Brady, Grimson
- 1981
(Show Context)
Citation Context |

11 | Model Based Recognition using Pruned Correspondence Search
- Breuel
- 1991
(Show Context)
Citation Context ...ire time that is high polynomial in the numbers of model and image features. Specifically, the required time is O(m 8n8), where m is the number of model features and n is the number of image features =-=[17, 21]-=-. Even if error is not carefully controlled, a system would still need to consider all minimally-sized hypotheses (sets of correspondences) between model and image features. For point features, there ... |

11 | Grouping Contours by Iterated Pairing Network - Shashua, Ullman - 1990 |

9 | Absolute Orientation from Uncertain Point Data: A Unified Approach
- Hel-Or, Werman
- 1991
(Show Context)
Citation Context ...tched to over-determine the pose. Some of this work analyzes the effect that errors in image features have on the accuracy of the resulting pose, including Kumar and Hanson [58] and Hel-Or and Werman =-=[45]-=-. The work of Hel-Or and Werman is particularly relevant to us, because they also consider how error propagates through the pose to the projections of unmatched feature points. Assuming Gaussian error... |

8 |
Shape completion,”
- Rutkowski
- 1979
(Show Context)
Citation Context |

7 |
Polynomial Time Object Recognition in the Presence of Clutter, Occlusion and Uncertainty
- Cass
- 1992
(Show Context)
Citation Context ...ire time that is high polynomial in the numbers of model and image features. Specifically, the required time is O(m 8n8), where m is the number of model features and n is the number of image features =-=[17, 21]-=-. Even if error is not carefully controlled, a system would still need to consider all minimally-sized hypotheses (sets of correspondences) between model and image features. For point features, there ... |

7 |
An automatic tube inspection system that finds cylinders in range data
- GRIMSON, LOZANO-PEREZ, et al.
- 1993
(Show Context)
Citation Context ...fication of defects in circuit boards, and the recognition and localization of objects on a conveyer belt (perhaps for subsequent manipulation by a robotic arm). As a specific example, Grimson et al. =-=[41]-=- recently reported a model-based system that accurately inspects three-dimensional, bent cylindrical sections, which are used in aircraft. The system performed alignment-based recognition and incorpor... |

6 |
Identification of Three-Dimensional Objects Using Range Information
- Reeves, Taylor
- 1989
(Show Context)
Citation Context ...s popular and has become so primarily for three reasons. First, the features used frequently are local, which means they are insensitive to occlusion. Non-local, region-based features such as moments =-=[27, 74]-=-, skeletons [85, 86], and geons [11] are difficult to obtain if regions can be partially occluded. Even though model-based representations may include such global features, in this work we are interes... |

5 | Shape encoding and subjective contours,” in
- Brady, Grimson, et al.
- 1980
(Show Context)
Citation Context |

5 |
The Integration of Figure Fragments into Representations of Planar Shape
- Elder, Zucker
- 1993
(Show Context)
Citation Context ... curve with a single large gap is preferred over the same curve with several small gaps of the same total size; this preference is inconsistent with the psychophysical experiments of Elder and Zucker =-=[28]-=-. In computing the saliency of a fragmented curve, gaps affect the total score in two ways (see Eq. 2.1). First, gap elements themselves do not contribute at all to the total score (since oj = 0 for v... |

5 | Uncertainty Estimates for Polyhedral Object Recognition - Ellis - 1989 |

5 |
Interactive Road Finding for Aerial Images
- Hu, Sakoda, et al.
- 1992
(Show Context)
Citation Context ...o prune the number of table entries that actually had to be generated. In our approach, a fixed graph is constructed on top of the image and all of the work is done in place. More recently, Hu et al. =-=[50]-=- used a shortest-paths approach to build a user-assisted system for finding roads in aerial images. Like us, Hu et al. searched a graph that is constructed on top of the image. However their graph is ... |

5 | Multi-Scale Vector-Ridge-Detection for Perceptual Organization Without Edges
- Subirana-Vilanova, Sung
- 1992
(Show Context)
Citation Context ...come so primarily for three reasons. First, the features used frequently are local, which means they are insensitive to occlusion. Non-local, region-based features such as moments [27, 74], skeletons =-=[85, 86]-=-, and geons [11] are difficult to obtain if regions can be partially occluded. Even though model-based representations may include such global features, in this work we are interested in features that... |

4 | The SRI Road Expert: Image-to-Database Correspondence - Bolles, Quam, et al. - 1978 |

4 | Fast Alignment Using Probabilistic Indexing
- Olson
- 1993
(Show Context)
Citation Context ...rtant is Pr[HIV] at distinguishing correct hypotheses. Its importance will depend on the effectiveness of the "probabilistic peaking effect," studied by Ben-arie [9] and applied to alignment by Olson =-=[70]-=-. The point of the probabilistic peaking effect is that, although all configurations of image triples can be produced by the model triple, some of these configurations may be much more likely. 5.7 Dis... |

4 |
The Shape of Subjective Contours
- Webb, Pervin
- 1984
(Show Context)
Citation Context |

3 | Robust and Efficient 3D Recognition by Alignment
- Alter
- 1992
(Show Context)
Citation Context ...overing 3D pose from three corresponding points. There are several solutions to this problem already, however, and instead we could apply one of them and then use the solution to compute 0 ands€ (see =-=[3]-=- for a review of the solutions). 4.3.2 First-Order Approximation Next we allow for error in one of the basis points, i 2 . The problem is to determine how fi• changes as a function of i2, with io and ... |

3 | High-accuracy Model Matching for Scenes Containing Man-Made Structures - Clark, Eckhardt, et al. - 1979 |

2 |
Perceptual Grouping for the Detection and Description
- Mohan, Nevatia
- 1988
(Show Context)
Citation Context ...tions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal [73, 53, 83, 72], or else incorporate substantial domain-specific information to constrain the possible groupings =-=[67, 98, 68]-=-. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that encourages long and straight curves. Using dynamic programming, they implemented a relaxation process on a ne... |

2 |
Perceptual Grouping for Scene Description and Segmentation
- Mohan, Nevatia
- 1992
(Show Context)
Citation Context ...tions [47, 26, 49, 55]. Other methods use greedy algorithms which need not be optimal [73, 53, 83, 72], or else incorporate substantial domain-specific information to constrain the possible groupings =-=[67, 98, 68]-=-. On the problem of saliency, Shashua and Ullman [82] proposed a measure for saliency that encourages long and straight curves. Using dynamic programming, they implemented a relaxation process on a ne... |

2 |
Curved Inertia Frames: Visual Attention and Perceptual Organization Using Convexity and Symmetry
- Subirana-Vilanova
- 1991
(Show Context)
Citation Context ...come so primarily for three reasons. First, the features used frequently are local, which means they are insensitive to occlusion. Non-local, region-based features such as moments [27, 74], skeletons =-=[85, 86]-=-, and geons [11] are difficult to obtain if regions can be partially occluded. Even though model-based representations may include such global features, in this work we are interested in features that... |

1 |
Riidiger von der Heydt, "A Computational Model of
- Heitger
- 1993
(Show Context)
Citation Context |