#### DMCA

## An efficient branch-and-bound algorithm for optimal human pose estimation

### Cached

### Download Links

Citations: | 20 - 3 self |

### Citations

8903 |
Probabilistic reasoning in intelligent systems: networks of plausible inference.
- Pearl
- 1988
(Show Context)
Citation Context ...d hi ∈ Hi, where |N | is the number of nodes in set N . Exact MAP inference over MRFs with large induced width is NP-hard [19]. Many approximate inference algorithms, such as loopy belief propagation =-=[12]-=- or generalized BP [30], have been proposed to obtain an efficient but non-optimal solution. Other approaches propose to relax the model into models which can be easily solved, such as star or tree mo... |

815 | Pictorial structures for object recognition - Felzenszwalb, Huttenlocher - 2005 |

395 |
Understanding belief propagation and its generalizations
- Yedidia, Freeman, et al.
- 2001
(Show Context)
Citation Context ...s the number of nodes in set N . Exact MAP inference over MRFs with large induced width is NP-hard [19]. Many approximate inference algorithms, such as loopy belief propagation [12] or generalized BP =-=[30]-=-, have been proposed to obtain an efficient but non-optimal solution. Other approaches propose to relax the model into models which can be easily solved, such as star or tree models [27, 18]. We follo... |

312 | An Automatic Method of Solving Discrete Programming Problems,"
- Land, Doig
- 1960
(Show Context)
Citation Context ... number of part location hypothesis) is large since its time complexity is proportional to the number of states to the power of the cluster size (typically ≥ 3). Methods based on Branch-andBound (BB) =-=[10]-=- search are used in Bayesian networks with a large number of random variables [11], but they become extremely inefficient when the number of states becomes larger (as in the human pose estimation prob... |

226 | Progressive search space reduction for human pose estimation.
- Ferrari, Marin-Jimenez, et al.
- 2008
(Show Context)
Citation Context ...nding a state-of-the-art tree model [16] to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such as Buffy =-=[7]-=- and Stickmen [5] datasets. Finally, our method can be used to exactly solve most of the inference problems on Stretchable Models [18] (which contains a few hundreds of variables) in just a few minute... |

211 | Pictorial structures revisited: People detection and articulated pose estimation,”
- Andriluka, Roth, et al.
- 2009
(Show Context)
Citation Context ...llance and gaming. Because speed is an important requirement in most of these applications, researchers have focused on approaches that put a premium on efficiency. Among them, tree-structured models =-=[6, 13, 5, 1, 15, 16, 22, 29]-=- (Fig. 1a) are commonly used. A tree structure typically captures only the most informative spatial relationships (i.e., kinematic constraints ) between pairs of parts since the location of one part i... |

160 | Fixing max-product: Convergent message passing algorithms for MAP LP-relaxations.
- Globerson, Jaakkola
- 2007
(Show Context)
Citation Context ... ∑ j∈Ni βji(h ∗ j , h ∗ i ) , (5) following the definition of hMAP and Eq. 3. There is a strong connection between our relaxed model and Linear Programming (LP) relaxation. Globerson and Jaakkola =-=[8]-=- derive a first-order Linear Programming (LP) relaxation, which intuitively can be interpreted as a mixture of star-models. This connection can be used to select B efficiently. For instance, B can be ... |

160 | Learning to parse images of articulated bodies,
- Ramanan
- 2006
(Show Context)
Citation Context ...llance and gaming. Because speed is an important requirement in most of these applications, researchers have focused on approaches that put a premium on efficiency. Among them, tree-structured models =-=[6, 13, 5, 1, 15, 16, 22, 29]-=- (Fig. 1a) are commonly used. A tree structure typically captures only the most informative spatial relationships (i.e., kinematic constraints ) between pairs of parts since the location of one part i... |

142 | Finding MAPs for belief networks is NP-hard. - Shimony - 1994 |

97 | Recovering human body configurations using pairwise constraints between parts,
- Ren, Berg, et al.
- 2005
(Show Context)
Citation Context ...n the image (over-counting evidence). To overcome these types of misclassification errors, more structured models such as the loopy graphical models (later referred as loopy model) have been proposed =-=[9, 20, 31, 14, 26, 28]-=- (Fig. 1b). By capturing interactions between a large number of pairs of parts, these methods are effective at improving pose estimation results at the expense of a significantly increased computation... |

89 | Better appearance models for pictorial structures. BMVC,
- Eichner, Ferrari, et al.
- 2009
(Show Context)
Citation Context ...the-art tree model [16] to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such as Buffy [7] and Stickmen =-=[5]-=- datasets. Finally, our method can be used to exactly solve most of the inference problems on Stretchable Models [18] (which contains a few hundreds of variables) in just a few minutes. 1. Introductio... |

82 | Measure locally, reason globally: Occlusion-sensitive articulated pose estimation.
- Sigal, Black
- 2006
(Show Context)
Citation Context ...n the image (over-counting evidence). To overcome these types of misclassification errors, more structured models such as the loopy graphical models (later referred as loopy model) have been proposed =-=[9, 20, 31, 14, 26, 28]-=- (Fig. 1b). By capturing interactions between a large number of pairs of parts, these methods are effective at improving pose estimation results at the expense of a significantly increased computation... |

78 |
Recursive star-tree parallel data structure.
- BERKMAN, VISHKIN
- 1993
(Show Context)
Citation Context ...) total time to speed up the query time computation, where E is the number of edges in the CRF. Notice that, in our earlier work [24], the computation is cast into a Range Maximum Query (RMQ) problem =-=[4]-=-. In this work, a simpler BMT data structure is used since the ranges (branches) are predefined according to the branching strategy in Algorithm 3. Most importantly, the second computation (argmaxhi∈H... |

72 | Cascaded models for articulated pose estimation
- Sapp, Toshev, et al.
- 2010
(Show Context)
Citation Context ...lem on loopy graphical models. We show that our method is empirically much faster (about 74 times) than the state-of-the-art exact inference algorithm [21]. By extending a state-of-the-art tree model =-=[16]-=- to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such as Buffy [7] and Stickmen [5] datasets. Finally, ... |

52 | Learning hierarchical poselets for human parsing.
- Wang, Tran, et al.
- 2011
(Show Context)
Citation Context ...n the image (over-counting evidence). To overcome these types of misclassification errors, more structured models such as the loopy graphical models (later referred as loopy model) have been proposed =-=[9, 20, 31, 14, 26, 28]-=- (Fig. 1b). By capturing interactions between a large number of pairs of parts, these methods are effective at improving pose estimation results at the expense of a significantly increased computation... |

52 |
Articulated pose estimation using flexible mixtures of parts.
- Yang, Ramanan
- 2011
(Show Context)
Citation Context ...llance and gaming. Because speed is an important requirement in most of these applications, researchers have focused on approaches that put a premium on efficiency. Among them, tree-structured models =-=[6, 13, 5, 1, 15, 16, 22, 29]-=- (Fig. 1a) are commonly used. A tree structure typically captures only the most informative spatial relationships (i.e., kinematic constraints ) between pairs of parts since the location of one part i... |

44 | A study of partsbased object class detection using complete graphs.
- Bergtholdt, Kappes, et al.
- 2010
(Show Context)
Citation Context ...arch only when the energy originated from the additional pair-wise relations is small (Fig.7 in [25]). This makes it hard for [25] to solve a model with many pair-wise interactions. Bergtholdt et al. =-=[3]-=- convert the inference problem over a fully connected model into a shortest path problem and propose an efficient A∗ search method for solving it. The main drawback of the A∗ search is that the branch... |

44 | Adaptive pose priors for pictorial structures, in:
- Sapp, Jordan, et al.
- 2010
(Show Context)
Citation Context |

44 | Parsing human motion with stretchable models.
- Sapp, Weiss, et al.
- 2011
(Show Context)
Citation Context ...dy parts (especially lower arms) on popular datasets such as Buffy [7] and Stickmen [5] datasets. Finally, our method can be used to exactly solve most of the inference problems on Stretchable Models =-=[18]-=- (which contains a few hundreds of variables) in just a few minutes. 1. Introduction Estimating the pose of humans (e.g., determining body part locations) from images and videos is a core problem in c... |

38 | Articulated part-based model for joint object detection and pose estimation - Sun, Savarese - 2011 |

36 | Improved Human Parsing with a Full Relational Model
- Tran, Forsyth
- 2010
(Show Context)
Citation Context |

35 | Multiple tree models for occlusion and spatial constraints in human pose estimation
- Wang, Mori
- 2008
(Show Context)
Citation Context ...eneralized BP [30], have been proposed to obtain an efficient but non-optimal solution. Other approaches propose to relax the model into models which can be easily solved, such as star or tree models =-=[27, 18]-=-. We follow this intuition and relax the model into a mixture of star-models (Sec 3.1). This is useful since the MAP solution of a mixture of starmodels can be obtained efficiently in time quadratical... |

31 |
Tightening LP relaxations for MAP using message-passing.
- Sontag, Meltzer, et al.
- 2008
(Show Context)
Citation Context ...and-bound to solve the human pose estimation problem on loopy graphical models. We show that our method is empirically much faster (about 74 times) than the state-of-the-art exact inference algorithm =-=[21]-=-. By extending a state-of-the-art tree model [16] to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such ... |

27 | Global pose estimation using non-tree models
- Jiang, Martin
(Show Context)
Citation Context |

27 |
Fast Globally Optimal 2d Human Detection with Loopy Graph
- Tian, Sclaroff
- 2010
(Show Context)
Citation Context ...e whole human body is easier to detect than the hands). A few works have been proposed to solve the MAP inference problem exactly for loopy models using efficient search algorithms. Tian and Sclaroff =-=[25]-=- propose an efficient BB algorithm for a tree model augmentedwith two additional pair-wise relations between left-right legs. The BB search is efficient since it only takes constant time to evaluate t... |

23 | Sidestepping intractable inference with structured ensemble cascades.
- Weiss, Sapp, et al.
- 2010
(Show Context)
Citation Context ...chniques to solve the joint estimation problem: A) exact inference on relaxed models, B) approximate inference on a full model (dual decomposition). Their experimental results on VideoPose2.0 dataset =-=[17]-=- show that (A) is both more efficient and accurate than (B). In this experiment, we first use message passing to select the best β in order to avoid having much looser upper bound. We show that our BB... |

18 | Tighter relaxations for map-mrf inference: A local primal-dual gap based separation algorithm
- Batra, Nowozin, et al.
- 2011
(Show Context)
Citation Context ...of the variable is split geometrically since we order the hypotheses so that geometrically nearby part hypotheses are also nearby in the ordered list (line 3 in Algorithm 3). Inspired by Batra et al. =-=[2]-=- and similarly to [24], we use a scoring function that we call Node-wise Primal Dual Gap (NPDG) as a cue to select the variable for branching (line 2 in Algorithm 3). Notice that the upper bound in Eq... |

17 | Best-first AND/OR search for graphical models
- Marinescu, Dechter
- 2007
(Show Context)
Citation Context ...b). By capturing interactions between a large number of pairs of parts, these methods are effective at improving pose estimation results at the expense of a significantly increased computational cost =-=[21, 11]-=-. For instance, methods based on cluster pursuit [21] become prohibitively slow when the number of states (i.e., number of part location hypothesis) is large since its time complexity is proportional ... |

17 |
margin and/or graph learning for parsing the human body
- Max
- 2008
(Show Context)
Citation Context |

9 | Efficient and exact MAP-MRF inference using branch and bound
- Sun, Telaprolu, et al.
- 2012
(Show Context)
Citation Context ...erence algorithm based on BB to solve the human pose estimation problem on loopy models, where the number of part location hypotheses is large. Our BB algorithm is built upon our earlier BB algorithm =-=[24]-=- for solving MAP inference on general MRF. Our contribution is two-fold: i) similarly to linear programming relaxation, a novel bound is obtained by relaxing the loopy model into a mixture of star-mod... |