Enhancing Exploration Efficiency using Uncertainty-Aware Information Prediction

Changseung Kim; Gaeun Yim; Heejung Shin; Hyondong Oh; SeungHwan Kim

REVIEW 3 major objections 6 minor 13 references

Enhancing Exploration Efficiency using Uncertainty-Aware Information Prediction

T0 review · 3 major / 6 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 robot explores unknown indoor environments faster when its expected information gain is computed from a Bayesian neural network's predicted map and the uncertainty of that prediction, evaluated as uniform fast Shannon mutual information.

desk verdict A coherent integration of MC-dropout occupancy prediction into FSMI, but the empirical speedup claim rests on 10 runs and none of the pairwise differences clears conventional significance. read the letter →

arxiv 2412.12825 v1 pith:XNZLXD77 submitted 2024-12-17 cs.RO

Seunghwan Kim , Heejung Shin , Gaeun Yim , Changseung Kim , Hyondong Oh This is my paper · ORCID

classification cs.RO

keywords autonomousexplorationoccupancygridmappredictionBayesianneuralnetworkmutualinformationfrontieruncertaintyquantificationMonteCarlodropoutmobilerobotics

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 frontier-exploration robot finishes mapping an unknown indoor environment faster when the expected information gain of each candidate viewpoint is evaluated with a learned prediction of the unseen part of the map together with a measure of how uncertain that prediction is. The proposal replaces the usual assumption that every unknown cell is exactly 50 percent occupied with probability values produced by a Bayesian neural network, and plugs those probabilities into a fast mutual-information computation. In GAZEBO simulations across four realistic indoor environments, the resulting metric, uniform FSMI with map prediction (P_Im), gave the shortest average completion time in every environment, with a smaller spread across trials than the prediction-based alternatives. The sympathetic reading is that using predictive uncertainty inside a probabilistic information metric, rather than as a separate variance score or a thresholded deterministic prediction, is what buys the efficiency gain.

What carries the argument

The load-bearing object is the Uniform FSMI with map prediction, denoted P_Im. Uniform FSMI is the fast Shannon mutual-information approximation that sums, over independent beams, the mutual information between an occupancy map and a future depth measurement under a uniform sensor-noise model; the proposed extension replaces the current-map probabilities $o_i$ with predicted occupancy probabilities $o^p_i = P(M^p_i = 1)$ obtained from a ResNet-18/U-Net Bayesian neural network with Monte Carlo dropout (ten stochastic forward passes per frontier), then evaluates $P I_m(V) = \sum_i MI(M; Z^i_p)$, where the per-beam term is $\sum_{k=0}^{n} P(e^p_k)(D_{k+H} - D_{k-H-1})/(2H+1)$. That formula is what turns a neural network's uncertainty from a separate variance map into an information-theoretic score.

What would settle it

Measure calibration of the predicted occupancy probabilities on held-out environments by binning predictions and comparing with actual occupancy frequencies; if the reliability diagram is far from the diagonal, the scores are biased. A second decisive check is to replace the learned predictions with the true map during evaluation: if P_Im no longer beats the best baseline, the benefit came from the prediction rather than the information metric.

Watch

Extended reading notes

Core claim

The central discovery the authors report is that when the occupancy probabilities used by uniform FSMI come from Monte-Carlo-dropout predictions of the unknown map, the exploration planner consistently outperforms six alternatives: deterministic volumetric gain with or without map prediction, variance-of-prediction metrics with either current-map or predicted-map raycasting, and the original mutual-information and volumetric-gain baselines. The claim is not only that the predicted map helps, but that the uncertainty channel matters: P_Im beats P_Iv (deterministic use of the same predicted map) in all four environments, and it beats the variance-only metrics P_Ivar1 and P_Ivar2. The authors attribute the gain to the fact that probabilistic treatment of future measurements in mutual information lets the planner weigh both expected coverage and confidence in the prediction.

Load-bearing premise

The whole benefit rests on treating the neural network's Monte-Carlo-dropout probabilities as honest probabilities about what is actually in unknown space; if those probabilities are systematically over- or underconfident, the information scores lose their meaning and the speedup could disappear.

Editorial extensions

If this is right

A frontier-exploration planner that uses P_Im should map the tested indoor environments faster than nearest-frontier, volumetric-gain, or original mutual-information baselines.
The advantage holds across all four environments, so the method is not tuned to a single layout.
Because P_Im also showed the smallest standard deviation among prediction-based methods in most environments, it can be expected to produce more consistent exploration times.
The predicted map alone (P_Iv) is not enough; the uncertainty-aware probabilistic treatment is what gives the additional speedup.

Reading between the lines

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

If the predictive probabilities remain well calibrated in new environments, the same uncertainty-aware FSMI scheme could transfer to 3D occupancy prediction or multi-robot coordination, where predicted occupancy in unmapped space could seed each robot's mutual-information computation.
The ten stochastic forward passes create a latency cost, so a model that achieves comparable calibration with fewer samples or with an analytic uncertainty estimate would make the approach more practical on embedded hardware.
A testable extension is to replace the uniform sensor-noise model with a learned or measured noise distribution; the FSMI structure permits this and would let the metric account for sensor-specific error.
The method's benefit likely scales with the quality of the prediction network, so in environments far from the training distribution P_Im may degrade toward baseline performance; a real-world trial would reveal whether the simulation advantage persists.

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

Read it if you work on exploration utility functions. The genuinely new piece is modest but real: prior learned-prediction work fed deterministic occupancy estimates into information gain, and here the U-Net's MC-dropout predictive probabilities are used directly as occupancy probabilities in the uniform FSMI ray model. That is a legitimate integration, not a new architecture or a new concept, and the paper is honest about that.

What it does well: the math in Section III.B.4 is coherent, Equations (20)-(22) connect cleanly to the FSMI derivation; the training setup is specified; the comparison includes six alternative metrics, including variance-based baselines; and four environments are tested. The authors also list their own limitations: no real-world validation, MC-dropout latency, and epistemic-only uncertainty. The citation pattern looks fine, and I see no evidence of fitted constants or outcome chasing.

Where it is soft: the empirical core. Table II reports 10 trials per environment with only means and standard deviations. The stress-test numbers are right: the P_Im vs P_Iv differences are 16.9s, 5.1s, 12.8s, and 13.2s across the four environments, with overlapping standard deviations and rough t-test p-values of 0.16, 0.41, 0.11, and 0.07. None of those reaches significance, even without multiple-comparison correction, and P_Iv is exactly the deterministic-control comparison the paper needs to beat. So 'consistently shows the shortest average exploration time' is true as a description of sample means, but 'superior performance' in the abstract is not statistically supported. I would not call this a fatal flaw--the effect may be real, and consistency across all four environments counts for something--but the claim outruns the evidence as written.

Second, the MC-dropout predictive distribution in Equation (15) is treated as calibrated occupancy probability in Equation (20), with no calibration analysis. Table I compounds the worry: recall is 0.554 and IoU is 0.518, so the predictions are not accurate enough that calibration can be assumed. This affects all downstream information scores, though the direction of bias is not obvious. Minor: Equation (8) looks like it has a typesetting or transcription error in the log term, and the authors should fix it.

Who this is for: people building frontier-based exploration systems with learned map priors. They will read this for the P_Im formulation and as a baseline reference. It deserves a serious referee--the integration is sensible, the math holds up, and the empirical weakness is fixable with more trials, significance tests, and a calibration check. I would send it to review and ask for those additions, not desk-reject it.

Referee Report

3 major / 6 minor

Summary. The paper presents an exploration method that combines occupancy grid map prediction with uncertainty quantification. A U-Net with MC dropout predicts occupancy probability in unknown areas, and the predictive distribution is plugged into a uniform fast Shannon mutual information (FSMI) formulation to score frontier viewpoints. The authors compare their method, P_Im, against six other information metrics in four GAZEBO environments with ten trials each, reporting that P_Im achieves the shortest average exploration time in every environment. The paper also reports network prediction accuracy and latency on a held-out test set.

Significance. If the reported advantage is real, integrating epistemic uncertainty from learned map prediction into an information-theoretic exploration objective is a sensible and potentially practically useful contribution. The mathematical derivation of uniform FSMI with predicted occupancy probabilities is coherent, and the manuscript is transparent about the main design choices (H, lambda, ns, dropout ratio) and about the simulated nature of the validation. The key weakness is that the central empirical claim rests on descriptive averages over ten trials, with overlapping standard deviations and no statistical testing; this is a load-bearing gap because P_Im is compared directly against its deterministic counterpart P_Iv, so the specific contribution of uncertainty-aware integration is exactly what is not shown to be statistically reliable.

major comments (3)

[Section IV-B, Table II] The claim that P_Im 'consistently shows the shortest average exploration time' is not statistically supported. With n=10 per environment, the differences between P_Im and the deterministic prediction baseline P_Iv are smaller than the reported standard deviations: Env 1: 157.54±18.30 vs 174.42±31.57 (two-sample t-test p≈0.16); Env 2: 155.68±9.76 vs 160.82±15.73 (p≈0.39); Env 3: 183.09±14.78 vs 195.92±18.71 (p≈0.11); Env 4: 77.64±9.02 vs 90.82±19.74 (p≈0.07). None of these pairwise comparisons reaches conventional significance, and no multiple-comparison correction or confidence intervals are reported. Since P_Iv is the deterministic counterpart of the proposed uncertainty-aware method, the core message of the paper depends on this comparison being reliable; the authors should report paired tests or effect sizes with confidence intervals, and ideally run additional trials or use variance-reduction techniques.
[Section III.A and III.B.4, Eqs. (15) and (20)] The predictive probabilities from MC dropout, obtained by averaging f(M_c,i; w_j) over ns=10 weight samples, are treated as calibrated occupancy probabilities o_p^i in the FSMI sensor model. The paper provides no calibration analysis, no reliability diagram, and no evidence that the predictive variance corresponds to the actual error distribution of the predicted map. If these probabilities are miscalibrated, the information scores in Eq. (21) are not faithful estimates of expected information gain, and the frontier ranking may be biased. The authors should add calibration evaluation for the predicted occupancy probabilities and, at minimum, discuss how miscalibration could affect the comparison between P_Im and the deterministic P_Iv baseline.
[Section IV-B, Fig. 5] The empirical evaluation aggregates only ten trials per environment, yet the paper draws a general conclusion about 'superior performance' from these averages. The simulation curves in Fig. 5 show mean exploration progress over time, but no error bands or per-trial variability are displayed. Reporting the distribution of exploration times, including scatter plots or confidence bands, and performing a statistical comparison across the four environments (e.g., a paired or mixed-effects analysis treating environment as a factor) would substantially strengthen the evidence. As it stands, the paper's central claim is not established beyond sampling noise.

minor comments (6)

[Eq. (8)] The typeset formula for f(δ,r) is ambiguous: it should be written as log((r+1)/(r+δ^{-1})) − log(δ/(rδ+1)) or equivalent, with parentheses clearly distinguishing numerator and denominator.
[Section III.B.2] The sentence 'In UPEN [8], the variance of neural network inference was used as an information metric' appears to cite the wrong reference: [8] is Kendall and Gal, while the UPEN method is [10] in the reference list.
[Eq. (9)] There is a typo: 'are are hyper parameters' should read 'are hyperparameters.'
[Section III.B.4, Eq. (22)] It is not explicitly stated whether the odds r_m used in the C_m terms of Eq. (22) are computed from the original OGM posterior or from the predicted occupancy probabilities o_p^i. This should be clarified, since it affects the interpretation of the mutual information being computed.
[Table II and Fig. 5] Please state explicitly in the table caption that the values are the mean and standard deviation of exploration time over 10 runs, and consider adding error bars or shaded confidence intervals to Fig. 5 so that trial-to-trial variability is visible.
[Section V] The conclusion mentions plans for real-world validation and aleatoric uncertainty, but the limitations paragraph could also acknowledge the lack of statistical significance testing in the current experiments, which is directly relevant to the strength of the reported empirical claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed P_Im metric is a composition of learned occupancy prediction and an existing mutual-information formula, evaluated against external simulator baselines.

full rationale

The paper's central claim is that integrating neural-network prediction uncertainty into uniform FSMI (P_Im) improves exploration efficiency. The derivation is self-contained: the learned prediction map P(M^p) is trained on labeled HouseExpo-derived GAZEBO data (Sec. III-A), and the information metric in Eq. (21)-(22) reuses the published Uniform FSMI formula [4] with the predicted occupancy probabilities o_i^p inserted into Eq. (20). No parameter is fitted to the reported exploration times, and the performance comparison in Table II is measured on an external simulator under a common frontier-exploration framework. The hyperparameters (dropout ratio, H, lambda) are declared and applied equally across methods, so they do not encode the target result. The paper does not rely on self-citations: the cited prior work is external, and no uniqueness theorem or author-imported ansatz is used to force the chosen metric. The acknowledged limitations (no real-world validation, MC-dropout latency, and neglect of aleatoric uncertainty) are correctness and generalization concerns, not circularity. Concerns about MC-dropout calibration or the statistical significance of the 10-trial averages are legitimate empirical criticisms, but they do not show that the derivation reduces to its inputs by construction. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends mainly on two categories of inputs: standard OGM and FSMI machinery, and the specific assumption that a neural network's predicted occupancy probabilities are reliable enough to drive sensor-model predictions. There are no newly invented physical entities or fit parameters that encode the target result.

free parameters (4)

H (sensor noise scaling factor) = unstated
This integer controls the width of the uniform sensor noise distribution in FSMI (Equation 13). It is chosen by hand and not reported in the experiments; it scales all mutual information values but does not appear to favor a specific metric.
lambda (distance cost weight) = 0.05
Used in the utility function (Equation 3) to trade off information against travel cost. Chosen by hand and applied identically to all methods.
ns (MC dropout samples) = 10
Number of stochastic forward passes in Equation (15). Affects the variance estimate and the smoothness of predicted probabilities; chosen without sensitivity analysis.
dropout ratio = 0.2
Dropout probability during training of the BNN. Directly controls the MC dropout uncertainty magnitude and is not tuned per environment.

assumptions (4)

domain assumption Occupancy grids are probabilistically independent (Equation 1).
Standard OGM assumption used in the Bayesian filter update and in the FSMI factorization over beams and grids.
standard math The uniform FSMI formula in Section II.D (from Zhang et al.) is correct and applicable.
The paper builds directly on the cited fast Shannon mutual information derivation without re-justifying it.
domain assumption MC dropout approximates a Bayesian posterior for the network (Equation 15).
Standard approximation from Gal and Ghahramani; treated as giving calibrated epistemic uncertainty, though the paper does not verify calibration.
ad hoc to paper Predicted occupancy probabilities o^p_i can substitute for true occupancy probabilities in the sensor model (Equation 20).
This is the core modeling choice of the paper. The network's softmax output is treated as the occupancy probability of unknown cells when computing beam hit distributions, with no empirical calibration check.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Exploration Efficiency using Uncertainty-Aware Information Prediction." pith.science (2026). https://pith.science/paper/XNZLXD77

@misc{pith2026241212825,
  author       = {Pith},
  title        = {Pith review of: Enhancing Exploration Efficiency using Uncertainty-Aware Information Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XNZLXD77}},
  note         = {Machine review of arXiv:2412.12825}
}

read the original abstract

Autonomous exploration is a crucial aspect of robotics, enabling robots to explore unknown environments and generate maps without prior knowledge. This paper proposes a method to enhance exploration efficiency by integrating neural network-based occupancy grid map prediction with uncertainty-aware Bayesian neural network. Uncertainty from neural network-based occupancy grid map prediction is probabilistically integrated into mutual information for exploration. To demonstrate the effectiveness of the proposed method, we conducted comparative simulations within a frontier exploration framework in a realistic simulator environment against various information metrics. The proposed method showed superior performance in terms of exploration efficiency.

Figures

Figures reproduced from arXiv: 2412.12825 by the authors.

**Figure 1.** Sample of the viewpoint generation. P(M|z1:t, x1:t) = Y 1≤i≤K P(Mi |z1:t, x1:t). (1) The update of the posterior probability distribution for each grid in the OGM can be performed using: P(Mi = 1|z1:t, x1:t) P(Mi = 0|z1:t, x1:t) = P(Mi = 1|z1:t−1, x1:t−1) P(Mi = 0|z1:t−1, x1:t−1) · P(Mi = 1|zt, xt) P(Mi = 0|zt, xt) . (2) In our paper, we will denote the posterior probability that the i-th grid is occupied as oi = P(… view at source ↗

**Figure 2.** Occupancy grid map prediction at the i-th frontier using a deep neural network. visibility is guaranteed from viewpoint V , which can be expressed as: P Ivar1(V ) = X z i∈z X Mp j ∈Mp∗ V ar(Mp j ), (18) where z i is obtained by performing raycasting on the current OGM, as in (4), and Mp∗ represents the set of grids in the prediction map that the beam z i passes through. 3) Variance of Prediction Map with Predicted R… view at source ↗

**Figure 3.** Illustrative results for each information metric. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Autonomous exploration simulation results. Each plot represents [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 4.** Figure 4: Simulation environments [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[1]

Receding horizon

A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon” next-best-view” planner for 3D exploration,” in Proc. IEEE Int. Conf. Robot. Autom. , 2016, pp. 1462–1468

work page 2016
[2]

On mutual information-based control of range sensing robots for mapping applications,

B. J. Julian, S. Karaman, and D. Rus, “On mutual information-based control of range sensing robots for mapping applications,” Int. J. Robot. Res. , vol. 33, no. 10, pp. 1375–1392, 2014

work page 2014
[3]

Information-theoretic mapping using Cauchy-Schwarz quadratic mutual information,

B. Charrow, S. Liu, V . Kumar, and N. Michael, “Information-theoretic mapping using Cauchy-Schwarz quadratic mutual information,” in Proc. IEEE Int. Conf. Robot. Autom. , 2015, pp. 4791–4798

work page 2015
[4]

FSMI: Fast computation of Shannon mutual information for information-theoretic mapping,

Z. Zhang, T. Henderson, S. Karaman, and V . Sze, “FSMI: Fast computation of Shannon mutual information for information-theoretic mapping,” Int. J. Robot. Res. , vol. 39, no. 9, pp. 1155–1177, 2020

work page 2020
[5]

Learned map prediction for enhanced mobile robot exploration,

R. Shrestha, F.-P. Tian, W. Feng, P. Tan, and R. Vaughan, “Learned map prediction for enhanced mobile robot exploration,” in Proc. IEEE Int. Conf. Robot. Autom. , 2019, pp. 1197–1204

work page 2019
[6]

Occupancy anticipation for efficient exploration and navigation,

S. K. Ramakrishnan, Z. Al-Halah, and K. Grauman, “Occupancy anticipation for efficient exploration and navigation,” in Comput Vis ECCV, 2020, pp. 400–418

work page 2020
[7]

Seer: Safe efficient exploration for aerial robots using learning to predict information gain,

Y . Tao, Y . Wu, B. Li, F. Cladera, A. Zhou, D. Thakur, and V . Kumar, “Seer: Safe efficient exploration for aerial robots using learning to predict information gain,” in Proc. IEEE Int. Conf. Robot. Autom. , 2023, pp. 1235–1241

work page 2023
[8]

What uncertainties do we need in Bayesian deep learning for computer vision?

A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?” Adv. Neural Inf. Process. Syst. , vol. 30, 2017

work page 2017

Show all 13 references

[9]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in Int. Conf. Mach. Learn., 2016, pp. 1050–1059

2016
[10]

Uncertainty-driven planner for exploration and navigation,

G. Georgakis, B. Bucher, A. Arapin, K. Schmeckpeper, N. Matni, and K. Daniilidis, “Uncertainty-driven planner for exploration and navigation,” in Proc. IEEE Int. Conf. Robot. Autom. , 2022, pp. 11 295– 11 302

2022
[11]

A frontier-based approach for autonomous explo- ration,

B. Yamauchi, “A frontier-based approach for autonomous explo- ration,” in Proc. IEEE Int. Symp. Comput. Intell. Robot. Autom , 1997, pp. 146–151

1997
[12]

Robot exploration with fast frontier detection: Theory and experiments,

M. Keidar and G. A. Kaminka, “Robot exploration with fast frontier detection: Theory and experiments,” in Proc. Int. Conf. Auton. Agents Multiagent Syst. , 2012, pp. 113–120

2012
[13]

Houseexpo: A large-scale 2d indoor layout dataset for learning-based algorithms on mobile robots,

L. Tingguang, H. Danny, L. Chenming, Z. Delong, W. Chaoqun, and M. Q.-H. Meng, “Houseexpo: A large-scale 2d indoor layout dataset for learning-based algorithms on mobile robots,” arXiv preprint arXiv:1903.09845, 2019

1903 arXiv

Pith tools

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

[1] [1]

Receding horizon

A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon” next-best-view” planner for 3D exploration,” in Proc. IEEE Int. Conf. Robot. Autom. , 2016, pp. 1462–1468

work page 2016

[2] [2]

On mutual information-based control of range sensing robots for mapping applications,

B. J. Julian, S. Karaman, and D. Rus, “On mutual information-based control of range sensing robots for mapping applications,” Int. J. Robot. Res. , vol. 33, no. 10, pp. 1375–1392, 2014

work page 2014

[3] [3]

Information-theoretic mapping using Cauchy-Schwarz quadratic mutual information,

B. Charrow, S. Liu, V . Kumar, and N. Michael, “Information-theoretic mapping using Cauchy-Schwarz quadratic mutual information,” in Proc. IEEE Int. Conf. Robot. Autom. , 2015, pp. 4791–4798

work page 2015

[4] [4]

FSMI: Fast computation of Shannon mutual information for information-theoretic mapping,

Z. Zhang, T. Henderson, S. Karaman, and V . Sze, “FSMI: Fast computation of Shannon mutual information for information-theoretic mapping,” Int. J. Robot. Res. , vol. 39, no. 9, pp. 1155–1177, 2020

work page 2020

[5] [5]

Learned map prediction for enhanced mobile robot exploration,

R. Shrestha, F.-P. Tian, W. Feng, P. Tan, and R. Vaughan, “Learned map prediction for enhanced mobile robot exploration,” in Proc. IEEE Int. Conf. Robot. Autom. , 2019, pp. 1197–1204

work page 2019

[6] [6]

Occupancy anticipation for efficient exploration and navigation,

S. K. Ramakrishnan, Z. Al-Halah, and K. Grauman, “Occupancy anticipation for efficient exploration and navigation,” in Comput Vis ECCV, 2020, pp. 400–418

work page 2020

[7] [7]

Seer: Safe efficient exploration for aerial robots using learning to predict information gain,

Y . Tao, Y . Wu, B. Li, F. Cladera, A. Zhou, D. Thakur, and V . Kumar, “Seer: Safe efficient exploration for aerial robots using learning to predict information gain,” in Proc. IEEE Int. Conf. Robot. Autom. , 2023, pp. 1235–1241

work page 2023

[8] [8]

What uncertainties do we need in Bayesian deep learning for computer vision?

A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?” Adv. Neural Inf. Process. Syst. , vol. 30, 2017

work page 2017