Pith. sign in

REVIEW 5 major objections 4 minor 59 references

Unifying Tree-Reweighted Belief Propagation and Mean Field for Tracking Extended Targets

T0 review · 5 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A factor graph split into a tree-reweighted BP region for data association and a mean-field region for state densities yields a closed-form extended-target tracker that avoids measurement clustering and gating.

desk verdict A real algorithmic contribution with a traceable derivation, but the empirical TRWBP-over-BP claim rests on rho_phi tuned on the test scenarios, so the headline performance advantage is not yet established. read the letter →

arxiv 2412.19036 v1 pith:2KQMEKGA submitted 2024-12-26 eess.SP

Weizhen Ma , Zhongliang Jing , Peng Dong , Henry Leung This is my paper · ORCID
classification eess.SP
keywords extendedtargettrackingfactorgraphtree-reweightedbeliefpropagationmeanfieldapproximationdataassociationgammaGaussianinverseWishartclosed-formrecursion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a single message-passing framework can handle the hard part of extended-target tracking, deciding which measurement came from which target, together with state estimation in one closed-form recursion. It splits the tracking factor graph into a belief-propagation (BP) region, where tree-reweighted BP resolves data association, and a mean-field (MF) region, where variational updates approximate the densities of measurement rate, kinematic state, and extent. For linear Gaussian target models with a gamma Gaussian inverse Wishart (GGIW) state density, the unified TRWBP-MF update closes in closed form, so no particles are needed and no measurement clustering or gating is performed. Simulation results on scenarios with ten and forty targets report lower GOSPA error and lower runtime than a max-sum association algorithm, a Poisson multi-Bernoulli mixture filter, and a particle-based BP-Poisson multi-Bernoulli filter.

What carries the argument

The load-bearing object is the unified TRWBP-MF message-passing scheme derived from Theorem 1, which identifies the stationary points of the region-based free energy Lagrangian with the fixed points of a set of message equations. The factor graph represents the joint posterior over potential-target existence, state, and two complementary association vectors; the graph is split into a BP region (factors $h_n$, $h_m$, $\phi_{nlm}$, $\phi_{mlo}$ that constrain data association) and an MF region (predictive and likelihood factors for the continuous state). TRWBP modifies BP by exponentiating factor messages with factor appearance probabilities $\rho_j$; the paper fixes $\rho_h=1$ and tunes a single scalar $\rho_\phi$ for the association-consistency factors. The MF messages give closed-form expectations against GGIW densities, and a fixed-point variational update, using a Gaussian for kinematic state, an inverse Wishart for extent, and auxiliary Gaussian measurement-source variables, closes the recursion.

What would settle it

Run TRWBP-MF and ordinary BP on a small scenario with, say, three targets and eight measurements, enumerate all valid association events exactly, and compare each algorithm's beliefs for the association variables against the exact posterior marginals; if the $\rho_\phi=0.15$ beliefs are not closer in KL divergence than the $\rho_\phi=1$ beliefs, the claimed improvement from tree-reweighting on this higher-order graph is not present.

Watch

Extended reading notes

Core claim

The central claim is that tree-reweighted belief propagation and mean-field approximation can be merged through a region-based free energy approximation into one inference engine on an extended-target factor graph. The BP region handles binary association variables via TRWBP message updates, which are claimed to converge more reliably than standard BP on the cyclic graph; the MF region computes approximate posterior densities for the continuous state. Under linear Gaussian dynamics and a GGIW prior over measurement rate, kinematic state, and extent, the algorithm yields a closed-form recursion: the measurement-rate belief is a mixture of gamma densities that is merged to a single gamma, the kinematic belief is Gaussian, and the extent belief is inverse Wishart, with a fixed-point variational step for their coupling. Because association is treated probabilistically, the tracker avoids hard measurement clustering and gating, and the paper's simulations report that it outperforms MSA, PMBM, and particle-based PMB-BP filters in accumulated GOSPA while running faster.

Load-bearing premise

The algorithm's claimed edge over ordinary belief propagation rests on the assumption that a single manually chosen weighting value on the data-association factors keeps the convergence and accuracy guarantees of tree-reweighted belief propagation, even though those guarantees are proven only for graphs whose factors connect two variables, while this graph's association factors connect more than two.

Editorial extensions

If this is right

  • Extended targets can be tracked without partitioning measurements into clusters or gating candidate associations, which is the failure mode the paper identifies for distance-based clustering when targets pass close together.
  • The state recursion is closed-form under the stated assumptions, so per-scan cost is polynomial in the number of targets and measurements rather than proportional to the number of particles.
  • Setting the association-factor appearance probability below one (the paper uses $\rho_\phi=0.15$) reduces localization and false-target GOSPA errors relative to plain BP ($\rho_\phi=1$) in both simulated scenarios.
  • The tracker jointly outputs existence probability, measurement rate, kinematic state, and extent for an unknown number of targets, which is what a detection-and-tracking system needs for lidar-like sensors.
  • Compared with MSA, PMBM, and particle-based PMB-BP, the paper's simulations show lower accumulated GOSPA and shorter runtime, with the largest speedup in the forty-target scenario.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the factor appearance probabilities were optimized per factor rather than fixed as a scalar $\rho_\phi$, the accuracy gains seen at $\rho_\phi=0.15$ might extend to a wider range of clutter and detection conditions; the paper's own sweeping result suggests the optimum is scenario-dependent.
  • Because the closed form lives in the MF region, non-linear or non-Gaussian dynamics could be absorbed there with sigma-point or particle approximations while keeping TRWBP for data association, giving a hybrid that retains the association benefits.
  • The paper avoids clustering and gating only for data association; track initialization still follows the message-censoring and measurement-clustering/reordering procedure of [18], so the claim of avoiding measurement clustering should be read with that qualification.
  • A natural testable extension is to compare TRWBP-MF against exact enumeration of association events in a small scenario; if the reweighted messages do not approximate the exact marginals better than BP, the constant-FAP assumption would need revisiting.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes TRWBP-MF, a message-passing algorithm for tracking extended targets that partitions the factor graph into a tree-reweighted belief propagation (TRWBP) region for data association and a mean-field (MF) region for the state densities. The authors derive a region-based free energy approximation, prove Theorem 1 relating stationary points to TRWBP-MF fixed points, and derive a closed-form recursion under linear Gaussian target models with gamma Gaussian inverse Wishart (GGIW) densities. The algorithm is evaluated on two simulated scenarios and compared with MSA, PMBM, and particle-based PMB-BP filters using the GOSPA metric and runtimes.

Significance. If the central claims are substantiated, the paper would offer a useful closed-form alternative to particle-based BP for extended target tracking, with a principled free-energy derivation rather than a purely heuristic message schedule. The strengths include a complete proof of Theorem 1 in the supplementary material, detailed derivations of the rescaled data-association messages, and explicit simulation settings that make the experiments reproducible. However, the theoretical and empirical support for the claimed advantage of TRWBP over standard BP is not yet established: the factor appearance probability is selected on the test scenarios, no argument connects the chosen FAPs to the guarantees of tree-reweighted BP for this factor graph, and several unquantified approximations enter the message derivations.

major comments (5)
  1. [Section V-B, Fig. 3, Table I] The central empirical claim that TRWBP (rho_phi = 0.15) outperforms BP (rho_phi = 1) is not established as stated, because the FAP rho_phi was selected by sweeping [0.1, 1] on the same two scenarios used for the final comparison. Since BP is exactly the rho_phi = 1 endpoint, and since the table reports only the selected rho_phi = 0.15 result without a held-out scenario or a validation protocol, the observed improvement may reflect selection on the test scenarios rather than a property of TRWBP. This is load-bearing for the Abstract's 'enhanced tracking performance' claim and for the conclusion that TRWBP is superior to ordinary BP.
  2. [Section II-C and Section IV-C] No argument connects the constant FAPs (rho_h = 1, rho_phi = 0.15) to the assumptions of the TRWBP theory in [37], which requires FAPs to lie in the spanning-tree polytope and is developed for pairwise factor graphs. The manuscript itself states in Section II-C that extending TRWBP to higher-order interactions 'is not straightforward,' yet the implementation sets rho_h = 1 for the higher-order detection factors h_n and h_m and reweights only the phi factors; no proof or numerical check shows that rho_phi = 0.15 is a valid FAP vector for the actual factor graph. Consequently, the claimed unique fixed-point and convergence properties from [37] do not transfer to this graph, weakening the theoretical motivation for the algorithm's advantage over BP.
  3. [Section IV-B, Eq. (29) and Supplementary Eq. (88)] The approximation (sE_n + R)^{-1} approx (sE_n)^{-1}, justified only by 'R is relatively small compared to sE_n,' is unquantified and is used both in the MF message updates and in the variational measurement update. With R = I_2 and s = 1/4 in the simulations, the accuracy of this approximation depends on the eigenvalues of E_n; no numerical check is provided. If the approximation is poor for targets with small extents, the closed-form updates in Eq. (62) and the likelihood in Eq. (59) may be biased, and this bias could affect both data association and state estimation.
  4. [Section IV-C, Eq. (35) and Eq. (37)] The derivation of the approximate message mBP->nl replaces mMF->n(x_n,r_n) * prod_l' mMF<-nl'(x_n,r_n) with the predicted density p+_n(x_n,r_n), and the analogous replacement is used for mBP->ml in Eq. (37). This is an additional approximation that is asserted without quantification or justification from the free-energy Lagrangian. Because these messages determine the association beliefs in Eqs. (47)-(48), the effect of this replacement on the association probabilities should be quantified or bounded before the closed-form recursion can be regarded as fully derived.
  5. [Abstract and Section V-B] The Abstract claims that the method 'avoids measurement clustering and gating,' but the implementation described in Section V-B initializes new tracks following [18], which 'consists of message censoring, and measurement clustering and reordering.' The method therefore avoids clustering only for the data-association update, not for birth initialization; this overstatement should be qualified in the Abstract and the introduction.
minor comments (4)
  1. [Section IV-C, Eqs. (32)-(38)] The message mBP,[iota]<-n(x_n,r_n) is used in Eq. (33) before it is defined in Eq. (45); reordering the presentation so that all messages are defined before first use would improve readability.
  2. [Table I caption] The caption says 'MEAN GOSPA ERROR' while also saying 'BOTH SUMMED OVER ALL THE TIME STEPS'; the table entries appear to be sums rather than means, so the caption should state which quantity is reported.
  3. [Fig. 3] No error bars or confidence intervals are shown in Fig. 3, even though results are averaged over 100 Monte Carlo runs; adding error bars would help the reader judge whether the differences between rho_phi values are significant.
  4. [Section V-B] The PMBM filter is run with a maximum of 20 assignments per partition, which is a restrictive truncation; the paper should justify that this truncation does not disadvantage PMBM relative to the proposed method, especially in Scenario 1 with high association uncertainty.

Circularity Check

1 steps flagged · score 3.0 of 10

TRWBP-MF derivation is self-contained, but the reported performance advantage rests on selecting ρϕ on the test scenarios, so part of the empirical claim is fitted rather than predicted.

  1. fitted input called prediction [Section V-B, Fig. 3 and Table I]
    "We first evaluate the performance of TRWBP-MF regarding the FAP ρϕ, which takes uniform values from 0.1 to 1, with an interval of 0.05. Fig. 3 shows the accumulated GOSPA errors for the two scenarios ... The minimum GOSPA errors are achieved at ρϕ = 0.15 and ρϕ = 0.2 for the first and second scenarios, respectively. ... Under pD = 0.95 and ρϕ = 0.15, Figs. 4 and 5 show the GOSPA errors ... Table I summarizes ... TRWBP-MF with ρϕ = 0.15 outperforms MSA, PMBM and PMB-BP in terms of GOSPA and localization error."

    The FAP ρϕ is not derived or set a priori; it is selected by sweeping the GOSPA metric over the same two scenarios that later serve as the evaluation benchmark in Table I. Therefore the reported 'TRWBP-MF (ρϕ = 0.15) outperforms ...' comparison is a statement about the best FAP found on the test data, not an independent prediction of TRWBP's advantage over BP (ρϕ = 1). The improvement from ρϕ = 1 to 0.15 is, by construction, a tuning result: no held-out scenario validates the chosen value, and the paper even notes that FAP optimization is future work. The theoretical derivation of the message-passing recursion is not circular, but this load-bearing empirical claim is partially a fit.

full rationale

The core derivation is not circular: Theorem 1 and the message recursions (20) and (21) are obtained by differentiating the region-based free energy Lagrangian (17)-(19), with the proof given in the supplementary material; there is no place where the message-passing equations are assumed as the conclusion. The self-citations in the paper are not load-bearing: [29] is used only as a comparison algorithm (MSA), and the TRWBP theoretical citations [37,38] are independent external work whose pairwise-graph limitation is explicitly acknowledged ('extending this approach to graphs with higher-order interactions is not straightforward'). The ρh = 1 simplification is an admitted implementation limitation rather than a circular reduction. The one substantive circular element is empirical: Section V-B chooses ρϕ by sweeping GOSPA on the same two scenarios that are then reported in Table I, and the claimed TRWBP-over-BP improvement is presented at the tuned value ρϕ = 0.15. This is a fitted input used as a prediction, since no validation scenario or a-priori FAP rule supports the chosen value; the magnitude of the reported advantage is therefore partially a tuning artifact rather than an independent test of TRWBP. The algorithmic derivation itself remains independent, so the paper is not fundamentally circular.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The algorithm rests on standard ETT modeling (Poisson measurement rate, GGIW prior, linear Gaussian kinematics) plus several approximations introduced for closed-form updates: small measurement noise relative to extent, constant FAPs with rho_h = 1, replacement of a full posterior message by the predicted density, and a fractional-power Gaussian likelihood. The key free parameter is the association FAP rho_phi, which is tuned on the evaluation scenarios.

free parameters (5)
  • rho_phi (FAP for association factors) = 0.15 (scenario 1 optimum; 0.2 for scenario 2; 0.15 used in Table I)
    Swept from 0.1 to 1 on the exact two scenarios used for evaluation, and the best value is used in the comparison against BP (rho_phi = 1). This tuning inflates the reported advantage of TRWBP over BP.
  • rho_h (FAP for detection factors) = 1
    Set to 1 in Section IV-C to simplify the derivation of m_BP_{->nl}; this is a hand-chosen simplification, not optimized.
  • eta = 1.1
    Heuristic factor for measurement rate prediction, adopted from [3,11,13].
  • tau = 50
    Heuristic factor for extent prediction, chosen by hand in the simulation setup.
  • lambda_n (new PT birth rate) = 0.01
    Poisson birth rate for new potential targets; scenario-dependent and chosen without a principled procedure.
assumptions (7)
  • domain assumption Inhomogeneous Poisson measurement model with gamma-distributed rate and GGIW joint state density
    Standard extended target model from [1,9,10]; the algorithm inherits this modeling assumption and is not valid outside it.
  • domain assumption Linear Gaussian kinematic model and Wishart extent evolution
    Required for the closed-form recursion; does not hold for maneuvering targets or non-elliptical extents.
  • ad hoc to paper R is small relative to sE_n
    Used in Section IV-B equations (26)-(29) to approximate E[ln|sE+R|] and E[(sE+R)^-1]; limits applicability to high-resolution sensors such as lidar.
  • ad hoc to paper Constant FAPs with rho_h = 1 preserve TRWBP benefits
    The TRWBP theory in [37] is for pairwise graphs; the paper sets constant FAPs and rho_h = 1 (Sections II-C and IV-C) without a convergence or optimality guarantee for the higher-order ETT graph.
  • ad hoc to paper Replacement of a full posterior message by the predicted density in m_BP_{->nl}
    Section IV-C states the term m_MF_{->n} prod m_MF_{<-nl} 'can be considered a reasonable approximation' to m_MF_{->n}, dropping the BP feedback that contains measurement-rate information.
  • ad hoc to paper Approximation (1 - pD + pD exp(-gamma)) approximately (1 - pD) in new PT message derivation
    Supplementary Section C.5, equation (49), pulls (1 - pD) out of the gamma integral, which is an approximation that affects the closed-form new-target messages.
  • ad hoc to paper Fractional power of a Gaussian treated as a Gaussian with scaled covariance
    Equation (59) raises a Gaussian to the power omega_m = l_n q_nl(m), which is generally not an integer; the reformulation with a normalizing constant is an approximation, though standard in variational ETT [50].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unifying Tree-Reweighted Belief Propagation and Mean Field for Tracking Extended Targets." pith.science (2026). https://pith.science/paper/2KQMEKGA

@misc{pith2026241219036,
  author       = {Pith},
  title        = {Pith review of: Unifying Tree-Reweighted Belief Propagation and Mean Field for Tracking Extended Targets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KQMEKGA}},
  note         = {Machine review of arXiv:2412.19036}
}
read the original abstract

This paper proposes a unified tree-reweighted belief propagation (BP) and mean field (MF) approach for scalable detection and tracking of extended targets within the framework of factor graph. The factor graph is partitioned into a BP region and an MF region so that the messages in each region are updated according to the corresponding region rules. The BP region exploits the tree-reweighted BP, which offers improved convergence than the standard BP for graphs with massive cycles, to resolve data association. The MF region approximates the posterior densities of the measurement rate, kinematic state and extent. For linear Gaussian target models and gamma Gaussian inverse Wishart distributed state density, the unified approach provides a closed-form recursion for the state density. Hence, the proposed algorithm is more efficient than particle-based BP algorithms for extended target tracking. This method also avoids measurement clustering and gating since it solves the data association problem in a probabilistic fashion. We compare the proposed approach with algorithms such as the Poisson multi-Bernoulli mixture filter and the BP-based Poisson multi-Bernoulli filter. Simulation results demonstrate that the proposed algorithm achieves enhanced tracking performance.

Figures

Figures reproduced from arXiv: 2412.19036 by the authors.

Figure 1
Figure 1. FG representation of the joint PDF in (8). We use abbreviations for factors and messages, removing their contained variables. For example, p + n represents p+ n (y n ), mMF →n represents mMF →n(y n ). C. The Unified TRWBP-MF message-passing Approach We first review the region-based free energy approximation, which is essential to the derivation of the unified TRWBP-MF approach. A region R = (IR,JR) in a FG consists … view at source ↗
Figure 2
Figure 2. True tracks for scenario 1 (left) and scenario 2 (righ [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 4
Figure 4. GOSPA errors and their components for scenario 1 give [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: GOSPA errors and their components for scenario 2 give [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages

  1. [18]

    Meyer and J

    F. Meyer and J. L. Williams, ``Scalable detection and tracking of geometric extended objects,'' IEEE Trans. Signal Process., vol. 69, pp. 6283--6298, Oct. 2021

    work page 2021

  2. [37]

    Wainwright, T

    M. Wainwright, T. Jaakkola, and A. Willsky, ``A new class of upper bounds on the log partition function,'' IEEE Trans. Inf. Theory, vol. 51, no. 7, pp. 2313--2335, Jul. 2005

    work page 2005

  3. [1]

    Gilholm and D

    K. Gilholm and D. Salmond, ``Spatial distribution model for tracking extended objects,'' IEE Proc. Radar, Sonar Navigat., vol. 152, no. 5, pp. 364--371, Oct. 2005

    work page 2005

  4. [2]

    J. W. Koch, ``Bayesian approach to extended object and cluster tracking using random matrices,'' IEEE Trans. Aerosp. Electron. Syst., vol. 44, no. 3, pp. 1042--1059, Jul. 2008

    work page 2008

  5. [3]

    Feldmann, D

    M. Feldmann, D. Fränken, and W. Koch, ``Tracking of extended objects and group targets using random matrices,'' IEEE Trans. Signal Process., vol. 59, no. 4, pp. 1409--1420, Apr. 2011

    work page 2011

  6. [4]

    Yang and M

    S. Yang and M. Baum, ``Tracking the orientation and axes lengths of an elliptical extended object,'' IEEE Trans. Signal Process., vol. 67, no. 18, pp. 4720--4729, Sep. 2019

    work page 2019

  7. [5]

    Tuncer and E

    B. Tuncer and E. Özkan, ``Random matrix based extended target tracking with orientation: A new model and inference,'' IEEE Trans. Signal Process., vol. 69, pp. 1910--1923, Mar. 2021

    work page 1910

  8. [6]

    Granström, C

    K. Granström, C. Lundquist, and U. Orguner, ``Tracking rectangular and elliptical extended targets using laser measurements,'' in Proc. 14th Int. Conf. Inf. Fusion, Jul. 2011, pp. 1--8

    work page 2011

Show all 59 references
  1. [7]

    Baum and U

    M. Baum and U. D. Hanebeck, ``Extended object tracking with random hypersurface models,'' IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 1, pp. 149--159, Jan. 2014

    2014

  2. [8]

    Wahlström and E

    N. Wahlström and E. Özkan, ``Extended target tracking using G aussian processes,'' IEEE Trans. Signal Process., vol. 63, no. 16, pp. 4165--4178, Aug. 2015

    2015

  3. [9]

    Granstr \"o m and U

    K. Granstr \"o m and U. Orguner, ``Estimation and maintenance of measurement rates for multiple extended target tracking,'' in Proc. 15th Int. Conf. Inf. Fusion, Singapore, Jul. 2012, pp. 2170--2176

    2012

  4. [10]

    Signal Process., vol

    ------, ``A PHD filter for tracking multiple extended targets using random matrices,'' IEEE Trans. Signal Process., vol. 60, no. 11, pp. 5657--5671, Nov. 2012

    2012

  5. [11]

    Lundquist, K

    C. Lundquist, K. Granström, and U. Orguner, ``An extended target CPHD filter and a gamma G aussian inverse W ishart implementation,'' IEEE J. Sel. Top. Signal Process., vol. 7, no. 3, pp. 472--483, Jun. 2013

    2013

  6. [12]

    Beard, S

    M. Beard, S. Reuter, K. Granstr \"o m, B.-T. Vo, B.-N. Vo, and A. Scheel, ``Multiple extended target tracking with labeled random finite sets,'' IEEE Trans. Signal Process., vol. 64, no. 7, pp. 1638--1653, Apr. 2016

    2016

  7. [13]

    Granstr \"o m, M

    K. Granstr \"o m, M. Fatemi, and L. Svensson, ``Poisson multi- B ernoulli mixture conjugate prior for multiple extended target filtering,'' IEEE Trans. Aerosp. Electron. Syst., vol. 56, no. 1, pp. 208--225, Feb. 2020

    2020

  8. [14]

    Y. Xia, K. Granstr \"o m, L. Svensson, M. Fatemi, \'A . F. García-Fern \'a ndez, and J. L. Williams, ``Poisson multi-bernoulli approximations for multiple extended object filtering,'' IEEE Trans. Aerosp. Electron. Syst., vol. 58, no. 2, pp. 890--906, Apr. 2022

    2022

  9. [15]

    Vivone and P

    G. Vivone and P. Braca, ``Joint probabilistic data association tracker for extended target tracking applied to X -band marine radar data,'' IEEE J. Ocean. Eng., vol. 41, no. 4, pp. 1007--1019, Oct. 2016

    2016

  10. [16]

    S. P. Coraluppi and C. A. Carthel, ``Multiple-hypothesis tracking for targets producing multiple measurements,'' IEEE Trans. Aerosp. Electron. Syst., vol. 54, no. 3, pp. 1485--1498, Jun. 2018

    2018

  11. [17]

    Granström, L

    K. Granström, L. Svensson, S. Reuter, Y. Xia, and M. Fatemi, ``Likelihood-based data association for extended object tracking using sampling methods,'' IEEE Trans. Intell. Veh., vol. 3, no. 1, pp. 30--45, Mar. 2018

    2018

  12. [19]

    Q. Li, R. Gan, J. Liang, and S. J. Godsill, ``An adaptive and scalable multi-object tracker based on the non-homogeneous P oisson process,'' IEEE Trans. Signal Process., vol. 71, pp. 105--120, Jan. 2023

    2023

  13. [20]

    S. Yang, L. M. Wolf, and M. Baum, ``Marginal association probabilities for multiple extended objects without enumeration of measurement partitions,'' in Proc. 23rd Int. Conf. Inf. Fusion, Rustenburg, South Africa, Jul. 2020, pp. 1--8

    2020

  14. [21]

    R. Gan, Q. Li, and S. J. Godsill, ``Variational tracking and redetection for closely-spaced objects in heavy clutter,'' IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 4, pp. 5286--5311, Aug. 2024

    2024

  15. [22]

    Kschischang, B

    F. Kschischang, B. Frey, and H.-A. Loeliger, ``Factor graphs and the sum-product algorithm,'' IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 498--519, Feb. 2001

    2001

  16. [23]

    Pearl, Probabilistic R easoning in I ntelligent S ystems: N etworks of P lausible I nference

    J. Pearl, Probabilistic R easoning in I ntelligent S ystems: N etworks of P lausible I nference . 1em plus 0.5em minus 0.4em San Mateo, CA, USA: Morgan Kaufmann, 1988

    1988

  17. [24]

    Williams and R

    J. Williams and R. Lau, ``Approximate evaluation of marginal association probabilities with belief propagation,'' IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 4, pp. 2942--2959, Oct. 2014

    2014

  18. [25]

    Meyer, P

    F. Meyer, P. Braca, P. Willett, and F. Hlawatsch, ``A scalable algorithm for tracking an unknown number of targets using multiple sensors,'' IEEE Trans. Signal Process., vol. 65, no. 13, pp. 3478--3493, Jul. 2017

    2017

  19. [26]

    Meyer, T

    F. Meyer, T. Kropfreiter, J. L. Williams, R. Lau, F. Hlawatsch, P. Braca, and M. Z. Win, ``Message passing algorithms for scalable multitarget tracking,'' Proc. IEEE, vol. 106, no. 2, pp. 221--259, Feb. 2018

    2018

  20. [27]

    Meyer and M

    F. Meyer and M. Z. Win, ``Scalable data association for extended object tracking,'' IEEE Trans. Signal Inf. Process. Netw., vol. 6, pp. 491--507, May 2020

    2020

  21. [28]

    Xia, \'A

    Y. Xia, \'A . F. Garc \'i a-Fern \'a ndez, F. Meyer, J. L. Williams, K. Granstr \"o m, and L. Svensson, ``Trajectory PMB filters for extended object tracking using belief propagation,'' IEEE Trans. Aerosp. Electron. Syst., pp. 1--23, Sep. 2023

    2023

  22. [29]

    W. Ma, Z. Jing, P. Dong, and H. Leung, ``Max sum based data associations for tracking point and extended targets,'' IEEE Trans. Aerosp. Electron. Syst., early access, doi:10.1109/TAES.2024.3482287 https://doi.org/10.1109/TAES.2024.3482287

    2024

  23. [30]

    C. M. Bishop and N. M. Nasrabadi, Pattern R ecognition and M achine L earning . 1em plus 0.5em minus 0.4em New York, NY, USA: Springer, 2006

    2006

  24. [31]

    Riegler, G

    E. Riegler, G. E. Kirkelund, C. N. Manchon, M.-A. Badiu, and B. H. Fleury, ``Merging belief propagation and the mean field approximation: A free energy approach,'' IEEE Trans. Inf. Theory, vol. 59, no. 1, pp. 588--602, Jan. 2013

    2013

  25. [32]

    Yedidia, W

    J. Yedidia, W. Freeman, and Y. Weiss, ``Constructing free-energy approximations and generalized belief propagation algorithms,'' IEEE Trans. Inf. Theory, vol. 51, no. 7, pp. 2282--2312, Jul. 2005

    2005

  26. [33]

    R. A. Lau and J. L. Williams, ``A structured mean field approach for existence-based multiple target tracking,'' in Proc. 19th Int. Conf. Inf. Fusion, Heidelberg, Germany, Jul. 2016, pp. 1111--1118

    2016

  27. [34]

    H. Lan, J. Ma, Z. Wang, Q. Pan, and X. Xu, ``A message passing approach for multiple maneuvering target tracking,'' Signal Process., vol. 174, p. 107621, Sep. 2020

    2020

  28. [35]

    H. Lan, Z. Wang, X. Bai, Q. Pan, and K. Lu, ``Measurement-level target tracking fusion for over-the-horizon radar network using message passing,'' IEEE Trans. Aerosp. Electron. Syst., vol. 57, no. 3, pp. 1600--1623, Jun. 2021

    2021

  29. [36]

    X. Bai, H. Lan, Z. Wang, Q. Pan, Y. Hao, and C. Li, ``Robust multitarget tracking in interference environments: A message-passing approach,'' IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 1, pp. 360--386, Feb. 2024

    2024

  30. [38]

    T. G. Roosta, M. J. Wainwright, and S. S. Sastry, ``Convergence analysis of reweighted sum-product algorithms,'' IEEE Trans. Signal Process., vol. 56, no. 9, pp. 4293--4305, Sep. 2008

    2008

  31. [39]

    T. P. Minka, ``Divergence measures and message passing,'' Microsoft Research, Cambridge, U.K., Tech. Rep. MSR-TR-2005-173, 2005

    2005

  32. [40]

    Chopra, ``On the spanning tree polyhedron,'' Oper

    S. Chopra, ``On the spanning tree polyhedron,'' Oper. Res. Lett., vol. 8, no. 1, pp. 25--29, Feb. 1989

    1989

  33. [41]

    Domke, ``Parameter learning with truncated message-passing,'' in CVPR 2011, Colorado Springs, CO, USA, Jun

    J. Domke, ``Parameter learning with truncated message-passing,'' in CVPR 2011, Colorado Springs, CO, USA, Jun. 2011, pp. 2937--2943

    2011

  34. [42]

    Wymeersch, F

    H. Wymeersch, F. Penna, and V. Savi \'c , ``Uniformly reweighted belief propagation: A factor graph approach,'' in 2011 IEEE Int. Symp. Inf. Theory, St. Petersburg, Russia, Jul./Aug. 2011, pp. 2000--2004

    2011

  35. [43]

    Wymeersch, F

    H. Wymeersch, F. Penna, and V. Savic, ``Uniformly reweighted belief propagation for estimation and detection in wireless networks,'' IEEE Trans. Wirel. Commun., vol. 11, no. 4, pp. 1587--1595, Apr. 2012

    2012

  36. [44]

    Winn and C

    J. Winn and C. M. Bishop, ``Variational message passing,'' J. Mach. Learn. Res, vol. 6, pp. 661--694, Apr. 2005

    2005

  37. [45]

    Dauwels, ``On variational message passing on factor graphs,'' in 2007 IEEE Int

    J. Dauwels, ``On variational message passing on factor graphs,'' in 2007 IEEE Int. Symp. Inf. Theory, Nice, France, Jun. 2007, pp. 2546--2550

    2007

  38. [46]

    Boyd and L

    S. Boyd and L. Vandenberghe, Convex optimization. 1em plus 0.5em minus 0.4em Cambridge, U.K.: Cambridge university press, 2004

    2004

  39. [47]

    K. B. Petersen and M. S. Pedersen, The M atrix C ookbook . 1em plus 0.5em minus 0.4em Kongens Lyngby, Denmark: Technical University of Denmark, 2012

    2012

  40. [48]

    A. K. Gupta and D. K. Nagar, Matrix V ariate D istributions . 1em plus 0.5em minus 0.4em New York, NY, USA: Chapman and Hall/CRC, 1999

    1999

  41. [49]

    Nguyen, ``An in depth introduction to variational B ayes note,'' Aug

    D. Nguyen, ``An in depth introduction to variational B ayes note,'' Aug. 2023, [ O nline]. A vailable: http://dx.doi.org/10.2139/ssrn.4541076 http://dx.doi.org/10.2139/ssrn.4541076

    2023 doi

  42. [50]

    Orguner, ``A variational measurement update for extended target tracking with random matrices,'' IEEE Trans

    U. Orguner, ``A variational measurement update for extended target tracking with random matrices,'' IEEE Trans. Signal Process., vol. 60, no. 7, pp. 3827--3834, Jul. 2012

    2012

  43. [51]

    D. G. Tzikas, A. C. Likas, and N. P. Galatsanos, ``The variational approximation for B ayesian inference,'' IEEE Signal Process. Mag., vol. 25, no. 6, pp. 131--146, Nov. 2008

    2008

  44. [52]

    A. S. Rahmathullah, \'A . F. Garc \'i a-Fern \'a ndez, and L. Svensson, ``Generalized optimal sub-pattern assignment metric,'' in Proc. 20th Int. Conf. Inf. Fusion, Xi'an, China, Jul. 2017, pp. 1--8

    2017

  45. [53]

    C. R. Givens and R. M. Shortt, ``A class of W asserstein metrics for probability distributions,'' Mich. Math. J., vol. 31, no. 2, pp. 231--240, Jan. 1984

    1984

  46. [54]

    Ester, H.-P

    M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, ``A density-based algorithm for discovering clusters in large spatial databases with noise,'' in Proc. 2nd Int. Conf. Knowl. Discovery Data Mining, Portland, OR, USA, Aug. 1996, pp. 226--231

    1996

  47. [55]

    W. Ma, Z. Jing, P. Dong, and H. Leung, ``Unifying tree-reweighted belief propagation and mean field for tracking extended targets''

  48. [56]

    Uniformly reweighted belief propagation: A factor graph approach,

    H. Wymeersch, F. Penna, and V. Savi´c, “Uniformly reweighted belief propagation: A factor graph approach,” in 2011 IEEE Int. Symp. Inf. Theory, St. Petersburg, Russia, Jul./Aug. 2011, pp. 2000–2004

    2011

  49. [57]

    Scalable data association for extended object tracking,

    F. Meyer and M. Z. Win, “Scalable data association for extended object tracking,” IEEE Trans. Signal Inf. Process. Netw., vol. 6, pp. 491–507, May 2020

    2020

  50. [58]

    Variational tracking and redetection for closely-spaced objects in heavy clutter,

    R. Gan, Q. Li, and S. J. Godsill, “Variational tracking and redetection for closely-spaced objects in heavy clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 4, pp. 5286–5311, Aug. 2024

    2024

  51. [59]

    C. M. Bishop and N. M. Nasrabadi, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2006

    2006

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.