Fast and Robust Initialization for Visual-Inertial SLAM

Carlos Campos; J. M. M. Montiel; Juan D. Tard\'os

REVIEW 4 major objections 4 minor 15 references

Fast and Robust Initialization for Visual-Inertial SLAM

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper claims that monocular visual-inertial SLAM initialization can be made fast and robust by adding two bundle-adjustment stages and two rejection tests to a closed-form solver.

desk verdict Solid engineering contribution that makes Martinelli-Kaiser initialization work in practice; the robustness claim is softer than the abstract implies because the key threshold is tuned on the same data. read the letter →

arxiv 1908.10653 v1 pith:22DY5EOT submitted 2019-08-28 cs.RO

Carlos Campos , J. M. M. Montiel , Juan D. Tardós This is my paper · ORCID

classification cs.RO

keywords visual-inertialSLAMinitializationmonocularscaleestimationbundleadjustmentobservabilityconsensustestrobustness

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 claims that the long, fragile startup of monocular visual-inertial SLAM can be replaced by a rapid joint initialization that works in one to two seconds. It starts from a closed-form solver for velocity, gravity direction, gyroscope bias, and scene depths, then sharpens it with two rounds of visual-inertial bundle adjustment. Two rejection tests—one on the smallest singular value of the first adjustment's Hessian, one on the fraction of additional tracked points that triangulate consistently—discard near-degenerate or spurious solutions. On the EuRoC V1 sequences the method reports consistent scale errors near 5% within two seconds, falling below 1% after ten more seconds of bundle adjustment. If this holds, the main obstacle to practical visual-inertial SLAM initialization is removed.

What carries the argument

The load-bearing object is the modified linear system $A(b_g)x = s(b_g,\alpha,\beta)$, built from feature tracks and IMU preintegration: for fixed gyro bias $b_g$ and a gravity direction parametrized by two angles $(\alpha,\beta)$, the feature depths and initial velocity $x$ are recovered by sparse least squares, while an outer optimization over $(b_g,\alpha,\beta)$ refines the nonlinear parameters. Two rounds of visual-inertial bundle adjustment follow; the observability test computes the smallest singular value of the first BA's Hessian and rejects the initialization if it falls below 0.1, and the consensus test triangulates all unused tracks and requires at least 90% to reproject within a chi-squared threshold. Preintegrated IMU deltas with first-order bias correction let the cost be evaluated without reintegrating inertial measurements on each bias update, which is what keeps the entire scheme real-time.

What would settle it

Run the full initialization on a dataset with a deliberately long near-pure-rotation segment followed by normal motion, and compare accepted versus rejected attempts against ground truth; if lowering the threshold to 0.01 admits many scale errors above 5%, or raising it to 1 rejects most valid starts, the threshold assumption fails. A simpler numeric check: sweep the threshold across 0.01, 0.1, and 1 on an unseen sequence and record false-acceptance and false-rejection rates.

Watch

Extended reading notes

Core claim

The paper's central claim is that the closed-form initialization it builds on, once generalized to partial tracks, given camera-IMU calibration, and made efficient with preintegration, is still too inaccurate and fragile to launch VI-SLAM; the missing ingredients are two stages of visual-inertial bundle adjustment and two cheap tests that reject bad solutions. The first test inspects the Hessian of the first BA and rejects any initialization whose smallest singular value falls below a threshold of 0.1, catching motions close to pure rotation or constant velocity that leave IMU variables weakly observable. The consensus test triangulates the unused tracked features, counts how many reproject within a chi-squared threshold across all observing frames, and accepts the initialization only if more than 90% agree, then runs a second, larger BA. With these additions the paper reports that on the EuRoC V1 sequences, accepted initializations have average scale error around 5%, in trajectories of 0.97 to 2.24 seconds, that full visual-inertial BA brings scale error below 1% after ten seconds, and that the whole check runs in under 215 ms of CPU time.

Load-bearing premise

The method depends on a single hand-chosen threshold, 0.1, applied to a numerical measure of how strongly the first optimization constrains all variables, to cleanly separate recoverable motions from near-degenerate ones; if that threshold does not transfer to new motions or sensors, the pipeline will either accept poor initializations or reject valid ones.

Editorial extensions

If this is right

Monocular VI-SLAM systems could start tracking in metric scale after 1 to 2 seconds instead of the roughly 15 seconds required by previous initialization pipelines.
Pure rotations and non-accelerated motions, which classically defeat inertial initialization, would be detected and rejected rather than silently poisoning the map.
A scale error near 5% at launch, tightening to under 1% after ten seconds of bundle adjustment, means maps become metrically usable almost immediately.
The same pipeline can run as a background thread while tracking continues, since initialization takes only 60 to 215 ms of CPU time.
In the reported experiments, initialization succeeds on the difficult V103 sequence where the previous pure-visual-then-inertial method failed.

Reading between the lines

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

The fixed threshold of 0.1 on the smallest singular value is a free parameter tuned on one dataset; on different camera-IMU noise levels the optimal cutoff likely shifts, and a noise-adaptive normalization of the Hessian's singular values would make the observability test portable.
The consensus test's 90% inlier bar may trade off recall in low-texture scenes; lowering it or weighting by track length could be tested against the same EuRoC sequences.
Because the observability test only needs the BA Hessian, the same singular-value criterion could be reused to detect degeneracy during ongoing VI-SLAM operation, not just at initialization.
The two-stage refinement suggests an analogous fast initialization for stereo-inertial setups where scale is directly observable, in which case the closed-form stage simplifies to velocity, gravity, and bias estimation.

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.

Desk Editor's Note

read the letter

Short version: this is a well-executed, practical VI-SLAM initialization paper, not a theoretical step change. It takes the Martinelli-Kaiser closed-form solver, extends it to partial tracks and camera-IMU extrinsics, adds preintegration, and wraps it in two BA stages with observability and consensus tests. That combination gets scale errors from roughly 30-150% down to around 5% on EuRoC V1, with convergence to under 1% after ten seconds of BA. I believe those numbers.

What is actually new: the generalization to partial tracks is the biggest enabler—MK assumed full tracks, which is unrealistic. The observability test via smallest singular value of the BA1 Hessian is a reasonable, cheap proxy for checking whether the motion excites the IMU states. The consensus test using unused tracks to reject bad initializations is also sensible. The preintegration rewrite is standard but makes the whole thing fast enough to run in real time. The experiments are exhaustive: hundreds of initialization attempts across three sequences, with a clear before/after table. The CPU comparison with and without preintegration is a nice touch.

Soft spots, in proportion. The tobs=0.1 threshold is the biggest one. It is picked from the same three sequences used for evaluation, no sensitivity analysis is given, and Figure 4 shows only one success and one failure. Since the reported 5% scale error is averaged only over solutions that pass the gate, the robustness claim depends on that scalar. I don't think this is fatal—the two-BA pipeline would likely still beat MK with any reasonable threshold—but it is a real gap that should be probed. The 'consistent initialization' phrasing is also optimistic: acceptance rates are 511/728, 101/395, and 71/336. That is still useful in practice because you can retry, but it is not 'consistently' in the everyday sense. Also, the scale-error metric in Table II is never explicitly defined. It matters, because the abstract's 5% number is that metric.

The citation pattern is clean: Martinelli, Kaiser, Forster preintegration, EuRoC, ORB-SLAM. No worries there. The math is plausible. This is not a paper that oversells under the hood.

Who is this for: anyone doing monocular VIO or VI-SLAM initialization. It is worth a serious referee; if I were handling it, I would send it out, and in revision I would require a definition of scale error and at least a small sensitivity study on tobs. The central argument holds up.

Referee Report

4 major / 4 minor

Summary. The paper proposes a fast and robust initialization method for monocular visual-inertial SLAM. It builds on the closed-form Martinelli-Kaiser solution, generalizing it to partial feature tracks and to the camera-IMU relative pose, and uses IMU preintegration for efficiency. The initial solution is refined by two rounds of visual-inertial bundle adjustment (BA1 and BA2), and two novel tests are applied: an observability test that rejects solutions when the smallest singular value of the BA1 Hessian falls below a threshold, and a consensus test that rejects solutions when the percentage of reprojection-inlier additional tracks is too low. Experiments on the three EuRoC V1 sequences report scale errors around 5% after the tests and below 1% after 10 seconds of visual-inertial BA, with initialization trajectories of 1-2 seconds and CPU times under 125 ms.

Significance. If the claims hold, the paper offers a practical speedup for monocular visual-inertial initialization, reducing the required trajectory duration from about 15 seconds in prior work to 1-2 seconds, with accuracy sufficient to launch a full VI-SLAM system. The derivation of the modified MK solution is clearly presented and the ablations show that the BA stages and the two tests improve accuracy, which is a useful engineering contribution. The experimental setup on the public EuRoC benchmark is reproducible in principle. However, the quantitative significance is currently weakened by the undefined scale-error metric, the single-environment evaluation with a threshold tuned on that same environment, the lack of sensitivity analysis for the main thresholds, and a speed claim that is not fully supported by the reported numbers. The core idea is promising and the defects are addressable, but the current evidence does not yet establish the robustness implied by the title.

major comments (4)

[Section IV (Tables II and IV)] The central quantitative claims are scale errors around 5% after the tests and below 1% after 10 seconds of BA, but the paper never defines the scale error metric. The tables report 'Scale error (%)' without specifying how the estimated scale is recovered from the initialization or how the percentage is computed. Please define the metric precisely, for example, the relative error of the scale factor obtained by aligning the estimated trajectory to ground truth over the initialization interval via a similarity transform, or the RMSE of the aligned trajectory divided by trajectory length. Without this definition, the reported numbers cannot be reproduced or interpreted, and this omission directly affects the evaluation of the main claim.
[Section III-A, Eq. (14), Table I, Figure 4] The observability gate uses a single hand-chosen threshold tobs = 0.1 on the smallest singular value of the BA1 Hessian. Because the Hessian is built from terms with different units and different information matrices (reprojection errors in pixels, IMU residuals, bias priors, Eq. (14)) and no bias-prior strengths are specified, the absolute value of its singular values is unit- and weighting-dependent. The threshold is selected on the same EuRoC V1 sequences used for evaluation, and no sensitivity analysis is provided. The paper should report the information matrices and bias-prior settings used in Eq. (14), show how the scale error and acceptance rate vary with tobs over at least a decade around 0.1, and validate the threshold on held-out sequences such as EuRoC V2 or MH. Without this, the claimed robustness of the method is not established.
[Abstract and Table II] The abstract claims that the method 'consistently initialize[s] in less than two seconds', but Table II reports an average initialization trajectory time of 2.235 seconds for V101 after filtering, so the statement is not supported by the data. Moreover, the reported 'Trajectory time' is the duration of the accepted initialization segment, not the time from system start to the first successful initialization. Given that 30% to 79% of attempts are rejected (Table II), the time-to-first-initialization could be substantially longer. Please clarify what the speed claim refers to and adjust the abstract accordingly, or provide the actual latency to first successful initialization.
[Section IV] The evaluation is confined to the three V1 sequences of EuRoC, the same environment used for tuning the observability and consensus thresholds. No results are given for the V2 or MH environments, even though EuRoC provides them. In its current form, the paper demonstrates that the method works well on V1, but the title and abstract claim robustness more broadly. At minimum, the paper should report results on V2 and MH or explicitly state that the robustness claim is limited to environments similar to EuRoC V1.

minor comments (4)

[Section II-B] The notation C_i = {C_1i, ..., C_ni} is confusing because the subindex i is used both for the feature index and the camera index. Consider using a different letter, for example k, for the camera index within a feature track.
[Figure 4] The y-axis is labeled 'Singular Value' but the axis label 'log10' appears at the top; it would be clearer to label the y-axis explicitly as 'log10 singular value' to avoid ambiguity about the scale.
[Section IV] The sentence 'This error converges to 1%' is imprecise because Table IV shows values below 1% at 10 seconds; changing it to 'is below 1%' would match the data.
[Table II] The caption says 'RMSE ATE is expressed in percentage over the length of the initialization trajectory', but the table also shows 'Scale error (%)' without a definition; please add a footnote defining the scale error.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation builds on the external Martinelli-Kaiser formulation and standard IMU preintegration, and the reported accuracy is measured against external EuRoC ground truth; the hand-tuned observability threshold is a generalization concern, not circularity.

full rationale

The paper's derivation chain is self-contained against external inputs. The core linear system (Eqs. 8-13) is constructed from the Martinelli-Kaiser formulation and standard IMU preintegration [7,8], and the accuracy claims are evaluated against EuRoC ground-truth trajectories (RMSE ATE and scale error), not against quantities fitted by the method. The observability test applies an SVD to the BA1 Hessian and compares the smallest singular value to a fixed threshold tobs=0.1, and the consensus test uses a chi-squared(95%) reprojection check; both are heuristic gates. The threshold tobs is selected on the same EuRoC V1 sequences used for evaluation, which is a generalization and overfitting concern rather than circularity: no equation defines the reported scale error as the threshold or as a fitted parameter, and the reported quantities are conditional statistics over accepted initializations, not the threshold itself. The only self-citations (ORB-SLAM [5],[13]) are used as an implementation base and baseline; they do not supply the derivation of the claimed improvements or the solvability conditions, which come from Martinelli [1] and Kaiser et al. [2]. No load-bearing step reduces by construction to its own inputs, so no circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim is tested on the EuRoC V1 dataset, which is external, so the derivation is not circular. However, the two test thresholds (tobs, tcons) and the BA hyperparameters are hand-tuned on the same sequences, which adds a mild overfitting caveat. No new physical entities are introduced.

free parameters (6)

tobs (observability test threshold) = 0.1
Hand-chosen threshold on the smallest Hessian singular value; no sensitivity analysis or theoretical justification in Section III-A.
tcons (consensus test inlier ratio) = 90%
Hand-chosen minimum inlier percentage for the consensus test in Section III-B; no sensitivity analysis.
m (features used in MK-solution) = 20
Number of tracked features used in the initial closed-form solution; chosen as a trade-off between speed and accuracy.
n (keyframes used in MK-solution) = 5
Number of keyframes used in the initial solution, chosen uniformly along time.
M (total number of tracks) = 200
Number of features tracked by ORB/KLT; chosen by the authors.
l (track-length test threshold) = 200 px
Minimum pixel length of tracks before an initialization attempt is launched.

assumptions (5)

domain assumption Accelerometer bias is negligible during the MK closed-form initial solution.
Section II-B states 'Neglecting accelerometer bias as in [2]'. The bias is only estimated later in BA1, so the initial solution could be biased for sensors with significant accelerometer bias.
standard math IMU preintegration first-order Jacobians from Forster et al. [8] are accurate enough for the initialization window.
Section II-B updates delta rotation and position terms using these Jacobians instead of reintegrating when biases change.
ad hoc to paper A threshold on the smallest singular value of the BA1 Hessian is a valid observability criterion.
Section III-A introduces this test with tobs=0.1, with only visual examples and no theoretical or statistical justification.
domain assumption The gravity magnitude is known and constant during initialization.
Section II-B models gravity as Exp(alpha,beta,0) gI with gI=(0,0,-g), treating g as a known constant.
domain assumption The camera-IMU extrinsic calibration [RBC|tBC] is known and accurate.
Section II-A uses the extrinsics in the feature equations; errors in calibration would propagate to all downstream estimates.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Fast and Robust Initialization for Visual-Inertial SLAM." pith.science (2026). https://pith.science/paper/22DY5EOT

@misc{pith2026190810653,
  author       = {Pith},
  title        = {Pith review of: Fast and Robust Initialization for Visual-Inertial SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22DY5EOT}},
  note         = {Machine review of arXiv:1908.10653}
}

read the original abstract

Visual-inertial SLAM (VI-SLAM) requires a good initial estimation of the initial velocity, orientation with respect to gravity and gyroscope and accelerometer biases. In this paper we build on the initialization method proposed by Martinelli and extended by Kaiser et al. , modifying it to be more general and efficient. We improve accuracy with several rounds of visual-inertial bundle adjustment, and robustify the method with novel observability and consensus tests, that discard erroneous solutions. Our results on the EuRoC dataset show that, while the original method produces scale errors up to 156%, our method is able to consistently initialize in less than two seconds with scale errors around 5%, which can be further reduced to less than 1% performing visual-inertial bundle adjustment after ten seconds.

Figures

Figures reproduced from arXiv: 1908.10653 by the authors.

**Figure 2.** Graph for the first visual-inertial BA. The body poses [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Singular values of the information matrix for a [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Initializations found along the EuRoC V101 trajec [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 12 canonical work pages

[1]

Closed-form solution of visual-inertial structure from motion,

A. Martinelli, “Closed-form solution of visual-inertial structure from motion,” International Journal of Computer Vision , vol. 106, no. 2, pp. 138–152, 2014

work page 2014
[2]

Simultaneous state initialization and gyroscope bias calibration in visual inertial aided navigation,

J. Kaiser, A. Martinelli, F. Fontana, and D. Scaramuzza, “Simultaneous state initialization and gyroscope bias calibration in visual inertial aided navigation,”IEEE Robotics and Automation Letters, vol. 2, no. 1, pp. 18–25, 2017

work page 2017
[3]

High-precision, consistent EKF-based visual-inertial odometry,

M. Li and A. I. Mourikis, “High-precision, consistent EKF-based visual-inertial odometry,” The International Journal of Robotics Re- search, vol. 32, no. 6, pp. 690–711, 2013

work page 2013
[4]

Keyframe-based visual–inertial odometry using nonlinear optimiza- tion,

S. Leutenegger, S. Lynen, M. Bosse, R. Siegwart, and P. Furgale, “Keyframe-based visual–inertial odometry using nonlinear optimiza- tion,” The International Journal of Robotics Research , vol. 34, no. 3, pp. 314–334, 2015

2015
[5]

Visual-inertial monocular SLAM with map reuse,

R. Mur-Artal and J. D. Tard ´os, “Visual-inertial monocular SLAM with map reuse,” IEEE Robotics and Automation Letters , vol. 2, no. 2, pp. 796–803, 2017

work page 2017
[6]

Robust initialization of monocular visual-inertial estimation on aerial robots,

T. Qin and S. Shen, “Robust initialization of monocular visual-inertial estimation on aerial robots,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2017, pp. 4225–4232

work page 2017
[7]

Visual-inertial-aided navigation for high- dynamic motion in built environments without initial conditions,

T. Lupton and S. Sukkarieh, “Visual-inertial-aided navigation for high- dynamic motion in built environments without initial conditions,” IEEE Transactions on Robotics , vol. 28, no. 1, pp. 61–76, 2012

work page 2012
[8]

IMU preinte- gration on manifold for efﬁcient visual-inertial maximum-a-posteriori estimation,

C. Forster, L. Carlone, F. Dellaert, and D. Scaramuzza, “IMU preinte- gration on manifold for efﬁcient visual-inertial maximum-a-posteriori estimation,” in Robotics: Science and Systems , 2015

work page 2015

Show all 15 references

[9]

The EuRoC micro aerial vehicle datasets,

M. Burri, J. Nikolic, P. Gohl, T. Schneider, J. Rehder, S. Omari, M. W. Achtelik, and R. Siegwart, “The EuRoC micro aerial vehicle datasets,” The International Journal of Robotics Research , vol. 35, no. 10, pp. 1157–1163, 2016

2016
[10]

ORB: An efﬁ- cient alternative to SIFT or SURF,

E. Rublee, V . Rabaud, K. Konolige, and G. Bradski, “ORB: An efﬁ- cient alternative to SIFT or SURF,” in IEEE International Conference on Computer Vision (ICCV) , 2011, pp. 2564–2571

2011
[11]

An iterative image registration technique with an application to stereo vision,

B. D. Lucas, T. Kanade, et al. , “An iterative image registration technique with an application to stereo vision,” in Int. Joint. Conf. on Artiﬁcial Intelligence (IJCAI) , 1981, pp. 674–679

1981
[12]

Szeliski, Computer vision: algorithms and applications

R. Szeliski, Computer vision: algorithms and applications . Springer Verlag, London, 2011

2011
[13]

ORB-SLAM: a versa- tile and accurate monocular SLAM system,

R. Mur-Artal, J. Montiel, and J. D. Tard ´os, “ORB-SLAM: a versa- tile and accurate monocular SLAM system,” IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147–1163, 2015

2015
[14]

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,” in IEEE Inter- national Conference on Robotics and Automation (ICRA) , 2011, pp. 3607–3613

2011
[15]

A benchmark for the evaluation of RGB-D SLAM systems,

J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of RGB-D SLAM systems,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2012, pp. 573–580

2012

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Closed-form solution of visual-inertial structure from motion,

A. Martinelli, “Closed-form solution of visual-inertial structure from motion,” International Journal of Computer Vision , vol. 106, no. 2, pp. 138–152, 2014

work page 2014

[2] [2]

Simultaneous state initialization and gyroscope bias calibration in visual inertial aided navigation,

J. Kaiser, A. Martinelli, F. Fontana, and D. Scaramuzza, “Simultaneous state initialization and gyroscope bias calibration in visual inertial aided navigation,”IEEE Robotics and Automation Letters, vol. 2, no. 1, pp. 18–25, 2017

work page 2017

[3] [3]

High-precision, consistent EKF-based visual-inertial odometry,

M. Li and A. I. Mourikis, “High-precision, consistent EKF-based visual-inertial odometry,” The International Journal of Robotics Re- search, vol. 32, no. 6, pp. 690–711, 2013

work page 2013

[4] [4]

Keyframe-based visual–inertial odometry using nonlinear optimiza- tion,

S. Leutenegger, S. Lynen, M. Bosse, R. Siegwart, and P. Furgale, “Keyframe-based visual–inertial odometry using nonlinear optimiza- tion,” The International Journal of Robotics Research , vol. 34, no. 3, pp. 314–334, 2015

2015

[5] [5]

Visual-inertial monocular SLAM with map reuse,

R. Mur-Artal and J. D. Tard ´os, “Visual-inertial monocular SLAM with map reuse,” IEEE Robotics and Automation Letters , vol. 2, no. 2, pp. 796–803, 2017

work page 2017

[6] [6]

Robust initialization of monocular visual-inertial estimation on aerial robots,

T. Qin and S. Shen, “Robust initialization of monocular visual-inertial estimation on aerial robots,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2017, pp. 4225–4232

work page 2017

[7] [7]

Visual-inertial-aided navigation for high- dynamic motion in built environments without initial conditions,

T. Lupton and S. Sukkarieh, “Visual-inertial-aided navigation for high- dynamic motion in built environments without initial conditions,” IEEE Transactions on Robotics , vol. 28, no. 1, pp. 61–76, 2012

work page 2012

[8] [8]

IMU preinte- gration on manifold for efﬁcient visual-inertial maximum-a-posteriori estimation,

C. Forster, L. Carlone, F. Dellaert, and D. Scaramuzza, “IMU preinte- gration on manifold for efﬁcient visual-inertial maximum-a-posteriori estimation,” in Robotics: Science and Systems , 2015

work page 2015