pith. sign in

arxiv: 2606.06250 · v1 · pith:OH7NK55Nnew · submitted 2026-06-04 · 💻 cs.RO

Breaking Time: A Fully Gaussian Framework for Distributed and Continuous-Time SLAM

Davide Ceriola , Simone Ferrari , Luca Di Giammarino , Leonardo Brizi , Giorgio Grisetti This is my paper

Pith reviewed 2026-06-28 01:03 UTC · model grok-4.3

classification 💻 cs.RO
keywords continuous-time SLAMGaussian Belief PropagationGaussian Process priorsdistributed optimizationasynchronous sensorsrolling shutter camerasmulti-camera SLAM
0
0 comments X

The pith

G-solver fuses Gaussian Belief Propagation with Gaussian Process priors to estimate continuous-time trajectories from asynchronous heterogeneous sensors in a distributed way.

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

The paper presents a method for continuous-time SLAM that represents trajectories probabilistically so that measurements from sensors with different timings can be fused without forcing them onto a common clock. It does this by pairing a Gaussian Process motion model, which supplies smooth interpolation between observations, with Gaussian Belief Propagation, which solves the resulting factor graph through local message exchanges. The approach is shown to extend directly to multiple unsynchronized cameras on both synthetic and real rolling-shutter data while keeping computation times comparable to prior continuous-time solvers. A sympathetic reader would care because many practical sensors produce data streams that are naturally staggered, and removing the need for explicit synchronization or post-processing steps could simplify the construction of multi-sensor systems.

Core claim

We introduce G-solver, a fully Gaussian and distributed framework that combines Gaussian Belief Propagation (GBP) with Gaussian Process (GP) motion priors for continuous-time trajectory estimation. Our GP model provides a probabilistic representation of the trajectory, enabling consistent interpolation and the use of data-driven hyperparameters, while GBP offers a scalable message-passing formulation well-suited for decentralized settings. The resulting solver naturally extends to multi-camera scenarios without specialized synchronization or engineering effort. We evaluate the approach on synthetic and real data, including rolling shutter and distributed multi-camera optimization, demonstrat

What carries the argument

G-solver, the pairing of Gaussian Belief Propagation message passing with a Gaussian Process motion prior that supplies both the trajectory representation and the factor-graph factors.

If this is right

  • Trajectories can be queried at any continuous time with a probability distribution rather than only at discrete measurement instants.
  • The same factor graph supports decentralized optimization across separate cameras or robots without a central clock.
  • Rolling-shutter distortion is absorbed directly into the continuous-time model instead of requiring separate undistortion preprocessing.
  • Hyperparameters of the motion prior can be learned from data inside the same inference procedure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same message-passing structure might allow incremental addition of new sensor modalities without redesigning the entire optimizer.
  • In multi-robot settings the distributed nature could reduce communication bandwidth compared with methods that require global synchronization.
  • Because the GP prior is data-driven, the framework may adapt motion statistics when the robot changes environment or speed without manual retuning.

Load-bearing premise

The Gaussian Process motion prior together with GBP message passing produces consistent interpolation and accurate estimates for asynchronous heterogeneous sensors without extra modeling or post-processing steps.

What would settle it

An experiment on real rolling-shutter or event-camera sequences in which the interpolated trajectory between measurement times shows larger reprojection errors or physically implausible motion than a synchronized discrete-time baseline.

Figures

Figures reproduced from arXiv: 2606.06250 by Davide Ceriola, Giorgio Grisetti, Leonardo Brizi, Luca Di Giammarino, Simone Ferrari.

Figure 1
Figure 1. Figure 1: Querying continuous-time mean and covariance. G-solver GP-based trajectory estimate on a torus motion sequence. Shaded regions depict the posterior covariance, which naturally expands in segments with fewer observations and contracts where measurements are available. The model supports querying both trajectory mean and uncertainty at arbitrary timestamps. is uncertainty handling: although spline uncertaint… view at source ↗
Figure 2
Figure 2. Figure 2: G-solver: Continuous-time inference via message passing. A factor graph view of our method: Gaussian Belief Propagation messages (arrows) propagate local information across the graph, while Gaussian Process priors model the tra￾jectory continuously in time. The light-blue region illustrates the GP uncertainty envelope. +1 +1 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Locally linear approximation between each pair of measurement times, ti and ti+1, where a constant velocity model is applied. This figure illustrates the relationships be￾tween the local pose variable ξi (t), and the global trajectory state {T(t), ϖ(t)}. The mean µ ∈ M lies on the manifold, while the covari￾ance Σ = Λ−1 ∈ R dim(TµM)×dim(TµM) is defined in the corresponding tangent space. The information ve… view at source ↗
Figure 4
Figure 4. Figure 4: Factor-to-Variable Message conditioning. The il￾lustration highlights the nodes (purple region) involved in computing the factor-to-variable message (light blue arrow). In Hyperion (right), the receiving variable is included in the conditioning set, whereas in our approach (left) it is excluded to avoid introducing bias into the computation. 0 5 10 15 20 25 # iteration 0 100 200 300 400 500 Etot Ours GN+GP… view at source ↗
Figure 5
Figure 5. Figure 5: Total energy over iterations. G-solver and Gauss￾Newton (GN)+GP total energy (Eq. (18)) while solving PGO in the Helix sequence, with initial perturbation and noise levels σ = 0.1, ηig = 1 (both [m] and [rad]). A G-solver iteration ends when every node has been updated and exchanged messages with all its neighbors. Here, the first sum encodes the energy of a generic set J of different measurement factors (… view at source ↗
Figure 7
Figure 7. Figure 7: Prior-based qualitative interpolation results. Pos￾terior continuous-time trajectory estimates for G-solver and Hyperion (B and Z spline) for the sphere sequence, under significant perturbation and noise levels with σ = 1, ηig = 1 (both [m] and [rad]). all approaches achieve comparable ATE and ARE RMSE values. In the lowest-noise regime, Z-Splines are slightly more accurate than G-solver, as their spline r… view at source ↗
Figure 10
Figure 10. Figure 10: Ablation on hyperparameters. ATE and ARE as a function of a scale factor s applied to the learned Q∗ C . The metrics are minimized between s = 1 and s = 10, but performance remains stable over a wide range. 0 5 10 15 20 25 Iteration 0 20 40 60 80 100 120 140 Mean Reprojection Error [px] Ours B-Spline Z-Spline [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: ChArUco reprojection error. Comparison of G￾solver and Hyperion for real-world BA using the mean of the reprojection error. s varies over several orders of magnitude. Performance peaks between s = 1 and s = 10. As expected, the optimal errors do not occur exactly at s = 1, since the accelerations along the torus trajectory are not constant. Nevertheless, performance degrades only moderately when scale s v… view at source ↗
Figure 13
Figure 13. Figure 13: Rolling shutter reprojection. Reprojection of opti￾mized 3D scene landmarks on KITTI 06 [34] under a 1 ms simulated readout time. Hyperion [26] results are shown in green and G-solver in blue. Nominal global shutter projections are shown in red. Our continuous-time GP model explicitly accounts for row-wise exposure, producing sharper and more temporally consistent reprojections than the spline-based Hy￾pe… view at source ↗
Figure 14
Figure 14. Figure 14: Multi-camera optimization scalability. (Left) The average computation time and memory per camera graph decrease as the number of camera segments increases, as the optimization load is distributed across multiple independent subgraphs. (Right) Estimation accuracy remains stable (ATE and ARE stay nearly constant), showing that distributing the optimization lowers computational cost without degrading accurac… view at source ↗
Figure 16
Figure 16. Figure 16: Prior-based experiments. Comparison between G￾solver and Hyperion using the sphere trajectory under varying measurement Gaussian noise standard deviations σ. The initial guess is perturbed with a standard deviation of ηig = 1 (applied to both positions [m] and orientations [rad]). All methods query the trajectory at a resolution 100× finer than the control-point spacing [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 18
Figure 18. Figure 18: PGO experiments. Comparison between G-solver and Hyperion using the sphere trajectory under varying mea￾surement Gaussian noise standard deviations σ. The initial guess is perturbed with a standard deviation of ηig = 1 (applied to both positions [m] and orientations [rad]). All methods query the trajectory at a resolution 100× finer than the control-point spacing [PITH_FULL_IMAGE:figures/full_fig_p012_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Rolling shutter reprojection. Reprojection of optimized 3D scene landmarks on KITTI 06 [34] under a 1 ms simulated readout time. Hyperion [26] results are shown in green and G-solver in blue. Nominal global shutter projections are shown in red [PITH_FULL_IMAGE:figures/full_fig_p013_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Rolling shutter reprojection. Reprojection of optimized 3D scene landmarks on KITTI 06 [34] under a 1 ms simulated readout time. Hyperion [26] results are shown in green and G-solver in blue. Nominal global shutter projections are shown in red [PITH_FULL_IMAGE:figures/full_fig_p014_20.png] view at source ↗
read the original abstract

Continuous-time SLAM provides a principled framework for fusing heterogeneous sensors while estimating smooth trajectories, and is particularly well-suited for handling heterogeneous, asynchronous sensor streams with non-uniform readout patterns, such as rolling shutter cameras, LiDAR scanners, radar sweeps, or event-based sensors. In this work, we introduce G-solver, a fully Gaussian and distributed framework that combines Gaussian Belief Propagation (GBP) with Gaussian Process (GP) motion priors for continuous-time trajectory estimation. Our GP model provides a probabilistic representation of the trajectory, enabling consistent interpolation and the use of data-driven hyperparameters, while GBP offers a scalable message-passing formulation well-suited for decentralized settings. The resulting solver naturally extends to multi-camera scenarios without specialized synchronization or engineering effort. We evaluate the approach on synthetic and real data, including rolling shutter and distributed multi-camera optimization, demonstrating accurate and stable estimation with runtimes comparable to existing continuous-time methods. An open-source implementation is released.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces G-solver, a fully Gaussian distributed framework for continuous-time SLAM that combines Gaussian Belief Propagation (GBP) with Gaussian Process (GP) motion priors. It claims this enables consistent interpolation of trajectories, handles asynchronous heterogeneous sensors (e.g., rolling shutter cameras), extends naturally to multi-camera scenarios without synchronization, and yields accurate stable estimation on synthetic and real data, with runtimes comparable to existing methods and an open-source implementation released.

Significance. If the central claims hold, the work would offer a scalable message-passing formulation for decentralized continuous-time trajectory estimation that integrates data-driven GP hyperparameters and handles multi-sensor asynchrony without extra engineering. The open-source release supports reproducibility, and the GBP+GP pairing is noted as compatible with prior literature.

major comments (1)
  1. [Abstract] Abstract: The claim of 'demonstrating accurate and stable estimation with runtimes comparable to existing continuous-time methods' on synthetic and real data (including rolling shutter and distributed multi-camera cases) is unsupported by any quantitative metrics, error bars, baseline comparisons, or derivation details; this directly undermines verification of the headline claims of consistency, stability, and practical performance.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their review and for highlighting the need for stronger support of the abstract claims. We address the comment below and will revise the manuscript to improve verifiability of the results.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim of 'demonstrating accurate and stable estimation with runtimes comparable to existing continuous-time methods' on synthetic and real data (including rolling shutter and distributed multi-camera cases) is unsupported by any quantitative metrics, error bars, baseline comparisons, or derivation details; this directly undermines verification of the headline claims of consistency, stability, and practical performance.

    Authors: We agree that the abstract would benefit from more explicit quantitative support to allow readers to immediately verify the performance claims. The experiments section of the manuscript presents results on synthetic and real datasets (including rolling-shutter and distributed multi-camera cases) with trajectory estimates, runtime measurements, and comparisons to existing continuous-time approaches. To directly address the concern, we will revise the abstract to incorporate key quantitative highlights (e.g., specific error values and runtime ratios relative to baselines) drawn from the experiments, and we will ensure the experiments section includes error bars, tabulated baseline comparisons, and clearer derivation details for the reported metrics. These changes will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper introduces G-solver as a new construction combining established GBP message passing with GP motion priors for continuous-time trajectory estimation. No load-bearing step reduces a claimed prediction or uniqueness result to a fitted input, self-definition, or self-citation chain by construction; the abstract and contributions frame the integration as an independent extension to distributed multi-camera settings without invoking prior author work to forbid alternatives or smuggle ansatzes. The framework is presented as self-contained against external benchmarks, with evaluation on synthetic and real data serving as independent validation rather than tautological confirmation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard properties of Gaussian processes for trajectory modeling and belief propagation for distributed inference; no explicit free parameters, new axioms, or invented entities are named in the abstract.

axioms (2)
  • domain assumption Gaussian Process motion priors provide a probabilistic representation enabling consistent interpolation
    Invoked when describing the GP component of the solver.
  • domain assumption Gaussian Belief Propagation offers a scalable message-passing formulation suitable for decentralized settings
    Invoked when describing the GBP component and its extension to multi-camera scenarios.

pith-pipeline@v0.9.1-grok · 5701 in / 1312 out tokens · 35766 ms · 2026-06-28T01:03:13.096890+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Orb-slam3: An accurate open-source library for visual, visual– inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard´os, “Orb-slam3: An accurate open-source library for visual, visual– inertial, and multimap slam,”IEEE Trans. on Robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  2. [2]

    Mast3r-slam: Real-time dense slam with 3d reconstruction priors,

    R. Murai, E. Dexheimer, and A. J. Davison, “Mast3r-slam: Real-time dense slam with 3d reconstruction priors,” inCVPR, pp. 16695–16705, 2025

    2025

  3. [3]

    Deep patch visual slam,

    L. Lipson, Z. Teed, and J. Deng, “Deep patch visual slam,” inECCV, pp. 424–440, Springer, 2024

    2024

  4. [4]

    Batch continuous-time trajectory estimation as exactly sparse gaussian process regression,

    T. D. Barfoot, C. H. Tong, and S. S ¨arkk¨a, “Batch continuous-time trajectory estimation as exactly sparse gaussian process regression,” in Proc. of Robotics: Science and Systems (RSS), 2014

    2014

  5. [5]

    Continuous-time state estimation methods in robotics: A survey,

    W. Talbot, J. Nubert, T. Tuna, C. Cadena, F. D ¨umbgen, J. Tordesillas, T. D. Barfoot, and M. Hutter, “Continuous-time state estimation methods in robotics: A survey,”IEEE Trans. on Robotics, 2025

    2025

  6. [6]

    Ortiz,Gaussian Belief Propagation for Real-time Decentralised In- ference

    J. Ortiz,Gaussian Belief Propagation for Real-time Decentralised In- ference. Imperial College London, 2023

    2023

  7. [7]

    Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image,

    E. Brachmann, F. Michel, A. Krull, M. Y . Yang, S. Gumhold, and C. Rother, “Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image,” inCVPR, pp. 3364–3372, 06 2016

    2016

  8. [8]

    Bundle adjustment - a modern synthesis,

    B. Triggs, P. F. McLauchlan, R. I. Hartley, and A. W. Fitzgibbon, “Bundle adjustment - a modern synthesis,” inICCV, ICCV ’99, pp. 298– 372, Springer-Verlag, 2000

    2000

  9. [9]

    Lsd-slam: Large-scale direct monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale direct monocular slam,” inECCV, pp. 834–849, 2014

    2014

  10. [10]

    Adaptive mobile robot navigation and mapping,

    H. J. S. Feder, J. J. Leonard, and C. M. Smith, “Adaptive mobile robot navigation and mapping,”Int. J. Robotic Res., vol. 18, no. 7, pp. 650– 668, 1999

    1999

  11. [11]

    Ceres Solver,

    S. Agarwal, K. Mierle, and T. C. S. Team, “Ceres Solver,” 10 2023

    2023

  12. [12]

    G2o: A general framework for graph optimization.,

    R. K ¨ummerle, G. Grisetti, H. Strasdat, K. Konolige, and W. Burgard, “G2o: A general framework for graph optimization.,” inProc. of the IEEE Intl. Conf. on Robotics Automation (ICRA), pp. 3607–3613, IEEE, 2011

    2011

  13. [13]

    Efficient derivative computation for cumulative b-splines on lie groups,

    C. Sommer, V . Usenko, D. Schubert, N. Demmel, and D. Cremers, “Efficient derivative computation for cumulative b-splines on lie groups,” inCVPR, pp. 11145–11153, 06 2020

    2020

  14. [14]

    Z-splines: moment conserving cardinal spline interpolation of compact support for arbitrarily spaced data,

    J. B. Sagredo and J. Tercero, “Z-splines: moment conserving cardinal spline interpolation of compact support for arbitrarily spaced data,” Seminar for Applied Mathematics (ETH Zurich) (SAM), 2003

    2003

  15. [15]

    Full steam ahead: Exactly sparse gaus- sian process regression for batch continuous-time trajectory estimation on se(3),

    S. Anderson and T. D. Barfoot, “Full steam ahead: Exactly sparse gaus- sian process regression for batch continuous-time trajectory estimation on se(3),” inProc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), p. 157–164, IEEE Press, 2015

    2015

  16. [16]

    A data- driven motion prior for continuous-time trajectory estimation on se (3),

    J. N. Wong, D. J. Yoon, A. P. Schoellig, and T. D. Barfoot, “A data- driven motion prior for continuous-time trajectory estimation on se (3),” IEEE Robotics and Automation Letters (RA-L), vol. 5, no. 2, pp. 1429– 1436, 2020

    2020

  17. [17]

    Vari- ational inference with parameter learning applied to vehicle trajectory estimation,

    J. N. Wong, D. J. Yoon, A. P. Schoellig, and T. D. Barfoot, “Vari- ational inference with parameter learning applied to vehicle trajectory estimation,”IEEE Robotics and Automation Letters (RA-L), vol. 5, no. 4, pp. 5291–5298, 2020

    2020

  18. [18]

    Continuous-time trajectory estimation: A comparative study between gaussian process and spline-based approaches,

    J. Johnson, J. Mangelson, T. Barfoot, and R. Beard, “Continuous-time trajectory estimation: A comparative study between gaussian process and spline-based approaches,” 2024

    2024

  19. [19]

    Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion,

    B. P. Duisterhof, L. Zust, P. Weinzaepfel, V . Leroy, Y . Cabon, and J. Revaud, “Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion,” inProc. of the International Conference on 3D Vision (3DV), pp. 1–10, IEEE, 2025

    2025

  20. [20]

    Data-efficient collaborative decentralized thermal-inertial odometry,

    V . Polizzi, R. Hewitt, J. Hidalgo-Carri ´o, J. Delaune, and D. Scaramuzza, “Data-efficient collaborative decentralized thermal-inertial odometry,” IEEE Robotics and Automation Letters (RA-L), vol. 7, no. 4, pp. 10681– 10688, 2022

    2022

  21. [21]

    C. M. Bishop,Pattern Recognition and Machine Learning (Information Science and Statistics). Berlin, Heidelberg: Springer-Verlag, 2006

    2006

  22. [22]

    Gaussian Belief Propagation: Theory and Aplication

    D. Bickson, “Gaussian belief propagation: Theory and aplication,” CoRR, vol. abs/0811.2518, 2008

    work page internal anchor Pith review Pith/arXiv arXiv 2008

  23. [23]

    Convergence analysis of distributed inference with vector-valued gaussian belief propagation,

    J. Du, S. Ma, Y .-C. Wu, S. Kar, and J. M. F. Moura, “Convergence analysis of distributed inference with vector-valued gaussian belief propagation,”Journal of Machine Learning Research, vol. 18, no. 172, pp. 1–38, 2018

    2018

  24. [24]

    Futuremapping: The computational structure of spatial ai systems,

    A. Davison, “Futuremapping: The computational structure of spatial ai systems,”CoRR, 03 2018

    2018

  25. [25]

    Futuremapping 2: Gaussian belief propagation for spatial ai,

    A. Davison and J. Ortiz, “Futuremapping 2: Gaussian belief propagation for spatial ai,” 10 2019

    2019

  26. [26]

    Hyperion – a fast, versatile symbolic gaussian belief propagation framework for continuous-time slam,

    D. Hug, I. Alzugaray, and M. Chli, “Hyperion – a fast, versatile symbolic gaussian belief propagation framework for continuous-time slam,” in ECCV, (Berlin, Heidelberg), p. 215–231, Springer-Verlag, 2024

    2024

  27. [27]

    Integrating generic sensor fusion algorithms with sound state representations through encap- sulation of manifolds,

    C. Hertzberg, R. Wagner, U. Frese, and L. Schr ¨oder, “Integrating generic sensor fusion algorithms with sound state representations through encap- sulation of manifolds,”Information Fusion, vol. 14, no. 1, pp. 57–77, 2013

    2013

  28. [28]

    Sparse gaussian processes on matrix lie groups: A unified framework for optimizing continuous-time trajectories,

    J. Dong, M. Mukadam, B. Boots, and F. Dellaert, “Sparse gaussian processes on matrix lie groups: A unified framework for optimizing continuous-time trajectories,” inProc. of the IEEE Intl. Conf. on Robotics Automation (ICRA), pp. 6497–6504, 2018

    2018

  29. [29]

    Batch nonlinear continuous-time trajectory estimation as exactly sparse gaussian process regression,

    S. Anderson, T. D. Barfoot, C. H. Tong, and S. S ¨arkk¨a, “Batch nonlinear continuous-time trajectory estimation as exactly sparse gaussian process regression,”Autonomous Robots, vol. 39, pp. 221–238, 2015

    2015

  30. [30]

    Message-passing for approximate map inference with latent variables,

    J. Jiang, P. Rai, and H. Daume, “Message-passing for approximate map inference with latent variables,” inNeurIPS(J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger, eds.), vol. 24, Curran Associates, Inc., 2011

    2011

  31. [31]

    T. D. Barfoot,State estimation for robotics. Cambridge University Press, 2024

    2024

  32. [32]

    A micro lie theory for state estimation in robotics,

    J. Sola, J. Deray, and D. Atchuthan, “A micro lie theory for state estimation in robotics,”arXiv preprint arXiv:1812.01537, 2018

    work page arXiv 2018

  33. [33]

    Ucoslam: Simultaneous localization and mapping by fusion of keypoints and squared planar markers,

    R. Munoz-Salinas and R. Medina-Carnicer, “Ucoslam: Simultaneous localization and mapping by fusion of keypoints and squared planar markers,”PR, vol. 101, p. 107193, 2020

    2020

  34. [34]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,”Int. J. Robotic Res., vol. 32, no. 11, pp. 1231–1237, 2013. 9 VI. SUPPLEMENTARYMATERIAL This section provides additional details to complement the main paper. We include a complete description of the inference components used in our framework, extended qualitative r...

    2013