Pictorial Structures for Object Recognition

Pedro F. FelzenszwalbDaniel P. Huttenlocher

article2004IJCV2,543 citations

Develops a computationally efficient statistical framework for part-based deformable object recognition, utilizing tree-structured models and generalized distance transforms to find globally optimal matches and sample posterior configurations in linear time.

Listen

The article addresses the challenge of recognizing generic classes of objects, such as faces and human bodies, in images. Traditional methods often struggle with deformable shapes and require manual model construction or inefficient search techniques, limiting their practicality for real-world applications where objects vary in pose and appearance.

The work evaluates an improved pictorial structure framework for part-based modeling. It sets out to develop efficient algorithms for matching models to images and for learning model parameters automatically from training examples.

The approach relies on tree-structured graphs with specific spring-like connections between parts. Experiments use labeled training images to learn appearance and spatial parameters via maximum likelihood estimation, followed by dynamic programming with generalized distance transforms for matching and posterior sampling for multiple hypotheses. Tests cover frontal faces from the Yale database and articulated human figures in binary images obtained by background subtraction.

The analysis shows that globally optimal matches can be found in time linear in the number of possible part locations rather than quadratic. Models learned from twenty face examples and ten body examples correctly locate objects in novel images, including cases with partial occlusion or noise. Sampling from the posterior distribution yields multiple high-quality hypotheses when the model is imprecise.

These results indicate that part-based deformable models can be made computationally practical and automatically constructed, reducing reliance on hand-tuned parameters or local search. The framework supports generic recognition tasks by jointly considering part appearance and geometric relations without early binary decisions about feature locations.

Further work is needed to handle more complex objects and larger pose spaces before broad deployment. The main limitations are the restriction to tree-structured models, discretization of part locations, and the need for post-sampling verification when parts overlap. Overall confidence in the reported efficiency and detection performance is high given the explicit experimental results.

Cornell University Massachusetts Institute of Technology efficient algorithm object appearance high resolution image Pictorial Structure Part-Based Model Energy Minimization

Felzenszwalb et al (2004).pdf

Paper: Comparing Images Using the Hausdorff Distance, Daniel P. Huttenlocher et al. (1993). Introduces the distance transform algorithms essential for computing optimal part configurations in pictorial structures in linear time.
Paper: Active Appearance Models, Timothy F. Cootes et al. (1998). Establishes statistical modeling of deformable shape and appearance variations that motivates flexible part-based object matching.

Paper: Object Detection with Discriminatively Trained Part-Based Models, Pedro F. Felzenszwalb et al. (2010). Directly builds on the pictorial structures framework by adding discriminatively trained mixtures of deformable part models with latent SVMs.
Paper: DeepPose: Human Pose Estimation via Deep Neural Networks, Alexander Toshev et al. (2014). Modernizes articulated human pose estimation by transitioning from tree-structured pictorial models to deep convolutional neural networks.
Paper: Convolutional Pose Machines, Shih-En Wei et al. (2016). Advances articulated part-based pose estimation beyond tree-structured graphical models by learning spatial part dependencies via sequential convolutional belief maps.
Paper: Realtime Multi-person 2D Pose Estimation Using Part Affinity Fields, Zhe Cao et al. (2016). Scales articulated part association to real-time multi-person scenes by replacing traditional tree-structured matching with nonparametric Part Affinity Fields.

Cover for Pictorial Structures for Object Recognition

Abstract

In this paper we present a computationally efficient framework for part-based modeling and recognition of objects. Our work is motivated by the pictorial structure models introduced by Fischler and Elschlager. The basic idea is to represent an object by a collection of parts arranged in a deformable configuration. The appearance of each part is modeled separately, and the deformable configuration is represented by spring-like connections between pairs of parts. These models allow for qualitative descriptions of visual appearance, and are suitable for generic recognition problems. We address the problem of using pictorial structure models to find instances of an object in an image as well as the problem of learning an object model from training examples, presenting efficient algorithms in both cases. We demonstrate the techniques by learning models that represent faces and human bodies and using the resulting models to locate the corresponding objects in novel images.

Topics

efficient algorithm object appearance high resolution image

Keywords

Pictorial Structure Part-Based Model Energy Minimization Tree-Structured Graph Maximum Likelihood MAP Estimate Mahalanobis Distance Gaussian Derivative Filter Sampling from Posterior Minimum Spanning Tree

1 Introduction
1.1 Pictorial Structures
1.2 Efficient Algorithms
1.3 Statistical Formulation
1.4 Related Work
2 Statistical Framework
3 Learning Model Parameters
3.1 Estimating the Appearance Parameters
3.2 Estimating the Dependencies
4 Matching Algorithms
4.1 Energy minimization or MAP Estimate
4.1.1 Efficient Minimization
4.1.2 Generalized Distance Transforms
4.2 Sampling from the Posterior
4.2.1 Computing the S functions
5 Iconic Models
5.1 Parts
5.2 Spatial Relations
5.3 Experiments
6 Articulated Models
6.1 Parts
6.2 Spatial Relations
6.3 Experiments
7 Summary
Acknowledgments
References

Knowls

Knowl 1 — Tree-Structured Pictorial Structure Model and Statistical Equivalence
model/method
A pictorial structure models an object as an undirected acyclic graph (tree) $G = (V, E)$ , where vertices $V = \{v_1, \dots, v_n\}$ correspond to $n$ object parts and edges $(v_i, v_j) \in E$ represent deformable connections between parts. An object configuration is defined by $L = (l_1, \dots, l_n)$ , where $l_i$ specifies the location (e.g., 2D position, scale, orientation) of part $v_i$ .

In the classical formulation, finding the optimal object configuration $L^*$ corresponds to minimizing the energy function:

$L^* = \arg\min_L \left( \sum_{i=1}^n m_i(l_i) + \sum_{(v_i, v_j) \in E} d_{ij}(l_i, l_j) \right)$

where $m_i(l_i)$ measures the visual mismatch of placing part $v_i$ at location $l_i$ , and $d_{ij}(l_i, l_j)$ measures the deformation cost between connected parts $v_i$ and $v_j$ .

In the statistical framework, parameterized by $\theta = (u, E, c)$ with part appearance parameters $u = \{u_1, \dots, u_n\}$ and connection parameters $c = \{c_{ij} \mid (v_i, v_j) \in E\}$ , the posterior distribution over configurations given an image $I$ is:

$p(L | I, \theta) \propto \left( \prod_{i=1}^n p(I | l_i, u_i) \right) \left( \prod_{(v_i, v_j) \in E} p(l_i, l_j | c_{ij}) \right)$

Assuming an uninformative prior over absolute part locations ( $p(l_i | \theta) = 1$ ), taking the negative logarithm demonstrates that Maximum A Posteriori (MAP) estimation is mathematically equivalent to the energy minimization problem, where $m_i(l_i) = -\log p(I | l_i, u_i)$ and $d_{ij}(l_i, l_j) = -\log p(l_i, l_j | c_{ij})$ .
Knowl 2 — Efficient MAP Energy Minimization via Generalized Distance Transforms
algorithm
For a tree-structured model $G = (V, E)$ rooted at an arbitrary vertex $v_r$ , the MAP configuration can be computed recursively. When deformation costs are constrained to be Mahalanobis distances between transformed locations:

$d_{ij}(l_i, l_j) = (T_{ij}(l_i) - T_{ji}(l_j))^T M_{ij}^{-1} (T_{ij}(l_i) - T_{ji}(l_j))$

with a diagonal matrix $M_{ij}$ , the dynamic programming step can be formulated as a generalized distance transform $\mathcal{D}_f(x) = \min_{y \in \mathcal{G}} (\rho(x, y) + f(y))$ , reducing the time complexity from $O(h^2 n)$ to $O(h' n)$ , where $h$ is the discrete number of part locations, $h'$ is the number of grid positions in the transformed space, and $n = |V|$ .
Input: Tree G = (V, E) rooted at v_r, unary costs m_i(l_i), mappings T_{ij}, T_{ji}, diagonal matrices M_{ij}
Output: Optimal configuration L^* = (l_1^*, ..., l_n^*)
for depth d from max_depth down to 1 do
    for each vertex v_j at depth d with parent v_i do
        Define f_j(y) over transformed grid G:
        if y in range(T_{ji}) then
            l_j = T_{ji}^{-1}(y)
            f_j(y) = m_j(l_j) + sum_{v_c in Children(v_j)} B_c(l_j)
        else
            f_j(y) = infinity
        end if
        Compute generalized distance transform under Mahalanobis distance M_{ij}:
        D_{f_j}(x) = min_{y in G} ((x - y)^T M_{ij}^{-1} (x - y) + f_j(y))
        Set B_j(l_i) = D_{f_j}(T_{ij}(l_i))
        Record best child location argmin B'_j(l_i) = T_{ji}^{-1}(arg min_{y in G} ((T_{ij}(l_i) - y)^T M_{ij}^{-1} (T_{ij}(l_i) - y) + f_j(y)))
    end for
end for
Compute optimal root location:
l_r^* = arg min_{l_r} (m_r(l_r) + sum_{v_c in Children(v_r)} B_c(l_r))
Trace back from root down to leaves:
for depth d from 0 up to max_depth - 1 do
    for each vertex v_i at depth d with known l_i^* do
        for each child v_j in Children(v_i) do
            l_j^* = B'_j(l_i^*)
        end for
    end for
end for
return L^* = (l_1^*, ..., l_n^*)
Knowl 3 — Exact Posterior Sampling Algorithm via Fast Gaussian Convolutions
algorithm
Exact samples of configurations $L \sim p(L | I, \theta)$ can be drawn in $O(h' n)$ time by computing tree marginals using Gaussian convolutions on the transformed coordinate grids.

When deformation priors follow a zero-mean Gaussian on transformed differences $p(l_i, l_j | c_{ij}) \propto \mathcal{N}(T_{ij}(l_i) - T_{ji}(l_j), 0, D_{ij})$ with diagonal covariance $D_{ij} = M_{ij}/2$ , the message $S_j(l_i)$ reduces to a separable Gaussian filtering operation: $S_j(l_i) \propto (F \otimes f_j)(T_{ij}(l_i))$ , where $F = \mathcal{N}(0, D_{ij})$ .
Input: Tree G = (V, E) rooted at v_r, likelihoods p(I | l_i, u_i), mappings T_{ij}, T_{ji}, diagonal covariances D_{ij}
Output: Sampled configuration L = (l_1, ..., l_n)
for depth d from max_depth down to 1 do
    for each vertex v_j at depth d with parent v_i do
        Define f_j(y) over transformed grid G:
        if y in range(T_{ji}) then
            l_j = T_{ji}^{-1}(y)
            f_j(y) = p(I | l_j, u_j) * prod_{v_c in Children(v_j)} S_c(l_j)
        else
            f_j(y) = 0
        end if
        Compute separable Gaussian convolution: G_j = F_{D_{ij}} conv f_j
        Set S_j(l_i) = G_j(T_{ij}(l_i))
    end for
end for
Compute root marginal distribution:
p(l_r | I, theta) proportional to p(I | l_r, u_r) * prod_{v_c in Children(v_r)} S_c(l_r)
Sample root location l_r from p(l_r | I, theta)
Sample children top-down from root to leaves:
for depth d from 0 up to max_depth - 1 do
    for each vertex v_i at depth d with sampled l_i do
        for each child v_j in Children(v_i) do
            Compute conditional distribution:
            p(l_j | l_i, I, theta) proportional to p(I | l_j, u_j) * p(l_i, l_j | c_{ij}) * prod_{v_c in Children(v_j)} S_c(l_j)
            Sample l_j from p(l_j | l_i, I, theta)
        end for
    end for
end for
return L = (l_1, ..., l_n)
Knowl 4 — Maximum Likelihood Estimation of Appearance, Deformations, and Tree Structure
algorithm
Given $m$ training images $\{I^1, \dots, I^m\}$ with annotated part configurations $\{L^1, \dots, L^m\}$ , all parameters $\theta = (u, E, c)$ of the pictorial structure model can be learned independently via Maximum Likelihood (ML) estimation.
Input: Training examples {(I^1, L^1), ..., (I^m, L^m)} with L^k = (l_1^k, ..., l_n^k)
Output: Optimal parameters theta^* = (u^*, E^*, c^*)
for each part v_i in V do
    Estimate part appearance parameters:
    u_i^* = arg max_{u_i} prod_{k=1}^m p(I^k | l_i^k, u_i)
end for
Set u^* = {u_1^*, ..., u_n^*}
for each pair of parts (v_i, v_j) with i != j do
    Estimate pairwise connection parameters:
    c_{ij}^* = arg max_{c_{ij}} prod_{k=1}^m p(l_i^k, l_j^k | c_{ij})
    Compute edge quality:
    q(v_i, v_j) = prod_{k=1}^m p(l_i^k, l_j^k | c_{ij}^*)
    Assign edge weight w(v_i, v_j) = -log q(v_i, v_j)
end for
Construct complete undirected graph G_complete = (V, E_complete) with edge weights w(v_i, v_j)
Compute the Minimum Spanning Tree (MST) E^* of G_complete using Kruskal's algorithm in O(n^2 log n) time
Set c^* = {c_{ij}^* | (v_i, v_j) in E^*}
return theta^* = (u^*, E^*, c^*)
Knowl 5 — Likelihood Annealing for Overlapping and Imprecise Part Models
model/method
When image likelihood is modeled as a product of independent part likelihoods $p(I|L, u) \propto \prod_{i=1}^n p(I|l_i, u_i)$ , overlapping parts over-count image evidence (e.g., explaining the same foreground pixels multiple times). This creates artificially high, false peaks in the posterior distribution $p(L|I, \theta)$ , causing MAP estimation to fail.

To counter this, posterior sampling is performed on an annealed (smoothed) likelihood distribution:

$p'(I|L, u) \propto p(I|L, u)^{1/T} = \prod_{i=1}^n p(I|l_i, u_i)^{1/T}$

where $T$ is a temperature parameter controlling the degree of smoothing (set to $T = 10$ for articulated body models). Annealing broadens posterior peaks, allowing the sampling procedure to generate diverse candidate configurations. Candidate samples are subsequently evaluated and re-ranked using an independent global metric that does not over-count evidence (such as the Chamfer distance between the proposed configuration shape and the binary input silhouette).
Knowl 6 — Iconic Part and Spatial Representation for Facial Feature Localization
model/method
In the iconic model formulation (e.g., for face localization), each part $v_i$ has a 2D pose $l_i = (x_i, y_i)$ .
1. Appearance Model: An iconic index $\alpha(l_i) \in \mathbb{R}^{27}$ is constructed by evaluating 9 Gaussian derivative filters at three standard deviations ( $\sigma_1 = 1, \sigma_2 = 2, \sigma_3 = 4$ ) and normalizing the response vector to achieve gain and bias invariance. The likelihood is modeled as a Gaussian with diagonal covariance $\Sigma_i$ and mean $\mu_i$ : $p(I | l_i, u_i) \propto \mathcal{N}(\alpha(l_i), \mu_i, \Sigma_i)$
2. Spatial Deformation Model: The relative spatial displacement $l_i - l_j$ between parts is parameterized by an ideal offset $s_{ij}$ and a full covariance matrix $\Sigma_{ij}$ : $p(l_i, l_j | c_{ij}) = \mathcal{N}(l_i - l_j, s_{ij}, \Sigma_{ij})$
3. Diagonalization for Distance Transforms: To satisfy the diagonal requirements of generalized distance transforms, the singular value decomposition $\Sigma_{ij} = U_{ij} D_{ij} U_{ij}^T$ is computed, yielding the coordinate transformations: $T_{ij}(l_i) = U_{ij}^T (l_i - s_{ij}), \quad T_{ji}(l_j) = U_{ij}^T l_j$ which transforms the deformation model into $p(l_i, l_j | c_{ij}) = \mathcal{N}(T_{ij}(l_i) - T_{ji}(l_j), 0, D_{ij})$ with diagonal covariance $D_{ij}$ .
Knowl 7 — Articulated Object Representation with von Mises Angular Relations
model/method
Articulated objects (such as human bodies) are modeled with parts parameterized in a 4D pose space $l_i = (x_i, y_i, s_i, \theta_i)$ , where $(x_i, y_i)$ is the center, $s_i \in [0, 1]$ represents foreshortening length scaling under scaled orthographic projection, and $\theta_i$ is orientation.
1. Appearance Model: For binary images with $t$ total pixels, likelihood is evaluated over rectangular parts (area $\text{area}_1$ , foreground count $\text{count}_1$ ) and surrounding border regions (area $\text{area}_2$ , foreground count $\text{count}_2$ ): $p(I | l_i, u_i) = q_1^{\text{count}_1} (1 - q_1)^{\text{area}_1 - \text{count}_1} q_2^{\text{count}_2} (1 - q_2)^{\text{area}_2 - \text{count}_2} 0.5^{t - \text{area}_1 - \text{area}_2}$ Foreground counts are evaluated efficiently across all translations via uniform filter box convolutions with 2-pixel dilation for $\text{count}_1$ and 2-pixel erosion for $\text{count}_2$ .
2. Spatial Joint Model: Connected parts $v_i, v_j$ meet at joint positions $(x'_i, y'_i)^T = (x_i, y_i)^T + s_i R_{\theta_i} (x_{ij}, y_{ij})^T$ and $(x'_j, y'_j)^T = (x_j, y_j)^T + s_j R_{\theta_j} (x_{ji}, y_{ji})^T$ , where $R_\theta$ is a 2D rotation matrix. Orientation difference follows a von Mises distribution $M(\theta_i - \theta_j, \theta_{ij}, k) \propto e^{k \cos(\theta_i - \theta_j - \theta_{ij})}$ .
3. Transform into 5D Diagonal Gaussian: Using the vector identity $\cos(\alpha - \beta) = -\frac{\|\vec{\alpha} - \vec{\beta}\|^2 - 2}{2}$ on unit vectors $\vec{\alpha} = [\cos \alpha, \sin \alpha]^T$ , the joint distribution is embedded into 5D transformed spaces: $T_{ij}(l_i) = (x'_i, y'_i, s_i, \cos(\theta_i + \theta_{ij}), \sin(\theta_i + \theta_{ij}))$ $T_{ji}(l_j) = (x'_j, y'_j, s_j, \cos(\theta_j), \sin(\theta_j))$ with diagonal covariance $D_{ij} = \text{diag}(\sigma_x^2, \sigma_y^2, \sigma_s^2, 1/k, 1/k)$ .
Knowl 8 — Empirical Face Localization and Robustness to Occlusion
empirical result
A 5-part iconic face model (representing the two eyes, nose, and two mouth corners) was trained on 20 images from the Yale face database with synthetic rotation and scale variations. Tree connectivity and spatial parameters were learned automatically using Maximum Likelihood estimation.

Matching using MAP estimation took less than 1 second per image on a desktop computer. When tested on partially occluded face images:
- When up to 2 of the 5 facial parts were occluded, the MAP estimate successfully and robustly localized all parts at reasonable configurations due to the learned tree deformation constraints.
- When 3 of the 5 parts were occluded, the global MAP match failed and placed parts incorrectly.
Knowl 9 — Empirical Articulated Human Body Pose Estimation via Sampling and Chamfer Re-ranking
empirical result
A 10-part articulated human body model (torso, head, two segments per arm, two segments per leg) was learned from 10 training images with manually labeled joint locations. The pose space for each part was discretized into a $70 \times 70 \times 10 \times 32$ grid over $(x, y, s, \theta)$ , yielding over $1.5 \times 10^6$ possible configurations per part.

Using annealed posterior sampling ( $T = 10$ ), 200 configuration samples were generated per input image and re-ranked using the Chamfer distance against background-subtracted binary silhouettes. Total execution time was approximately 1 minute per image on a desktop computer. The method successfully recovered accurate 2D articulated poses despite substantial clutter, limb articulation, and corrupted binary inputs with large missing foreground regions.

Coverage note — The 9-part iconic model for facial gaze direction was omitted as a separate knowl because its formulation is an identical application of the iconic model knowl, differing only in part count.

References

1.A.A. Amini, T.E. Weymouth, and R.C. Jain. Using dynamic programming for solving variational problems in vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(9):855–867, September 1990.
2.Y. Amit and D. Geman. A computational model for visual selection. Neural Computation, 11(7):1691–1715, October 1999.
3.N.J. Ayache and O.D. Faugeras. Hyper: A new approach for the recognition and positioning of two-dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence, 8(1):44–54, January 1986.
4.J.O. Berger. Statistical Decision Theory and Bayesian Analysis. Springer-Verlag, 1985.
5.G. Borgefors. Distance transformations in digital images. Computer Vision, Graphics, and Image Processing, 34(3):344–371, June 1986.
6.G. Borgefors. Hierarchical chamfer matching: A parametric edge matching algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 10(6):849–865, November 1988.
7.Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1222–1239, November 2001.
8.C. Bregler and J. Malik. Tracking people with twists and exponential maps. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8–15, 1998.
9.M.C. Burl and P. Perona. Recognition of planar object classes. In IEEE Conference on Computer Vision and Pattern Recognition, pages 223–230, 1996.
10.M.C. Burl, M. Weber, and P. Perona. A probabilistic approach to object recognition using local photometry and global geometry. In European Conference on Computer Vision, pages II:628–641, 1998.
11.C.K. Chow and C.N. Liu. Approximating discrete probability distributions with dependence trees. IEEE Transactions on Information Theory, 14(3):462–467, May 1968.
12.T.H. Cormen, C.E. Leiserson, and Rivest R.L. Introduction to algorithms. MIT Press and McGraw-Hill, 1996.
13.S.J. Dickinson, I. Biederman, A.P. Pentland, J.O. Eklundh, R. Bergevin, and R.C. Munck-Fairwood. The use of geons for generic 3-d object recognition. In International Joint Conference on Artificial Intelligence, pages 1693–1699, 1993.
14.P.F. Felzenszwalb and D.P. Huttenlocher. Efficient matching of pictorial structures. In IEEE Conference on Computer Vision and Pattern Recognition, pages II:66–73, 2000.
15.M.A. Fischler and R.C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381–395, June 1981.
16.M.A. Fischler and R.A. Elschlager. The representation and matching of pictorial structures. IEEE Transactions on Computer, 22(1):67–92, January 1973.
17.W.T. Freeman and E.H. Adelson. The design and use of steerable filters. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(9):891–906, September 1991.
18.Y. Gdalyahu and D. Weinshall. Flexible syntactic matching of curves and its application to automatic hierarchical classification of silhouettes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(12):1312–1328, December 1999.
19.S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 6(6):721–741, November 1984.
20.W.E.L. Grimson and T. Lozano-Perez. Localizing overlapping parts by searching the interpretation tree. IEEE Transactions on Pattern Analysis and Machine Intelligence, 9(4):469–482, July 1987.
21.E.J. Gumbel, J.A. Greenwood, and D. Durand. The circular normal distribution: Theory and tables. Journal of the American Statistical Association, 48:131–152, March 1953.
22.D.P. Huttenlocher, G.A. Klanderman, and W.J. Rucklidge. Comparing images using the hausdorff distance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(9):850–863, September 1993.
23.D.P. Huttenlocher and S. Ullman. Recognizing solid objects by alignment with an image. International Journal of Computer Vision, 5(2):195–212, November 1990.
24.S. Ioffe and D.A. Forsyth. Probabilistic methods for finding people. International Journal of Computer Vision, 43(1):45–68, June 2001.
25.H. Ishikawa and D. Geiger. Segmentation by grouping junctions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 125–131, 1998.
26.S.X. Ju, M.J. Black, and Y. Yacoob. Cardboard people: A parameterized model of articulated motion. In International Conference on Automatic Face and Gesture Recognition, pages 38–44, 1996.
27.A.V. Karzanov. Quick algorithm for determining the distances from the points of the given subset of an integer lattice to the points of its complement. Cybernetics and System Analysis, pages 177–181, April-May 1992. Translation from the Russian by Julia Komissarchik.
28.Y. Lamdan, J.T. Schwartz, and H.J. Wolfson. Affine invariant model-based object recognition. IEEE Transactions on Robotics and Automation, 6(5):578–589, 1990.
29.B. Moghaddam and A.P. Pentland. Probabilistic visual learning for object representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(7):696–710, July 1997.
30.H. Murase and S.K. Nayar. Visual learning and recognition of 3-d objects from appearance. International Journal of Computer Vision, 14(1):5–24, January 1995.
31.J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, 1988.
32.A.P. Pentland. Recognition by parts. In IEEE International Conference on Computer Vision, pages 612–620, 1987.
33.L. Rabiner and B. Juang. Fundamentals of Speech Recognition. Prentice Hall, 1993.
34.D. Ramanan and D.A. Forsyth. Finding and tracking people from the bottom up. In IEEE Conference on Computer Vision and Pattern Recognition, pages II: 467–474, 2003.
35.R.P.N. Rao and D.H. Ballard. An active vision architecture based on iconic representations. Artificial Intelligence, 78(1-2):461–505, October 1995.
36.E. Rivlin, S.J. Dickinson, and A. Rosenfeld. Recognition by functional parts. Computer Vision and Image Understanding, 62(2):164–176, September 1995.
37.L.G. Roberts. Machine perception of 3-d solids. In Optical and Electro-optical Information Processing, pages 159–197, 1965.
38.W. Rucklidge. Efficient Visual Recognition Using the Hausdorff Distance. Springer-Verlag, 1996. LNCS 1173.
39.T.B. Sebastian, P.N. Klein, and B.B. Kimia. Recognition of shapes by editing shock graphs. In IEEE International Conference on Computer Vision, pages I: 755–762, 2001.
40.M. Turk and A.P. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):71–96, 1991.
41.W.M. Wells, III. Efficient synthesis of Gaussian filters by cascaded uniform filters. IEEE Transactions on Pattern Analysis and Machine Intelligence, 8(2):234–239, March 1986.

Citation

MLA

Felzenszwalb, P. F., and D. P. Huttenlocher. “Pictorial Structures for Object Recognition”. International Journal of Computer Vision, vol. 61, no. 1, 2005, pp. 55–79, https://doi.org/10.1023/B:VISI.0000042934.15159.49.

APA

Felzenszwalb, P. F., & Huttenlocher, D. P. (2005). Pictorial Structures for Object Recognition. International Journal of Computer Vision, 61(1), 55–79. https://doi.org/10.1023/B:VISI.0000042934.15159.49

Chicago

Felzenszwalb, P. F., and D. P. Huttenlocher. 2005. “Pictorial Structures for Object Recognition”. International Journal of Computer Vision 61 (1): 55–79. https://doi.org/10.1023/B:VISI.0000042934.15159.49.

Harvard

Felzenszwalb, P.F. and Huttenlocher, D.P. (2005) “Pictorial Structures for Object Recognition”, International Journal of Computer Vision, 61(1), pp. 55–79. Available at: https://doi.org/10.1023/B:VISI.0000042934.15159.49.

Vancouver

1. Felzenszwalb PF, Huttenlocher DP (2005) Pictorial Structures for Object Recognition. International Journal of Computer Vision 61:55–79

BibTeX

@article{Felzenszwalb_2005, title={Pictorial Structures for Object Recognition}, volume={61}, ISSN={1573-1405}, url={http://dx.doi.org/10.1023/B:VISI.0000042934.15159.49}, DOI={10.1023/b:visi.0000042934.15159.49}, number={1}, journal={International Journal of Computer Vision}, publisher={Springer Science and Business Media LLC}, author={Felzenszwalb, Pedro F. and Huttenlocher, Daniel P.}, year={2005}, month=Jan, pages={55–79} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Pictorial Structures for Object Recognition

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Tree-Structured Pictorial Structure Model and Statistical Equivalence

Knowl 2 — Efficient MAP Energy Minimization via Generalized Distance Transforms

Knowl 3 — Exact Posterior Sampling Algorithm via Fast Gaussian Convolutions

Knowl 4 — Maximum Likelihood Estimation of Appearance, Deformations, and Tree Structure

Knowl 5 — Likelihood Annealing for Overlapping and Imprecise Part Models

Knowl 6 — Iconic Part and Spatial Representation for Facial Feature Localization

Knowl 7 — Articulated Object Representation with von Mises Angular Relations

Knowl 8 — Empirical Face Localization and Robustness to Occlusion

Knowl 9 — Empirical Articulated Human Body Pose Estimation via Sampling and Chamfer Re-ranking

References

Citation

BibTeX

Access the Paper