Distributional Uncertainty for Out-of-Distribution Detection

DaeUng Jo; Jeonghyo Song; Jinyoung Kim; Kimin Yun; Youngjoon Yoo

REVIEW 5 major objections 5 minor 17 references

Distributional Uncertainty for Out-of-Distribution Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · 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 replacing RPL's fixed free-energy target with a variance map from a Beta posterior network improves out-of-distribution detection in semantic segmentation.

desk verdict The paper's central mechanism — the variance-based loss that supposedly replaces RPL's free-energy target — is identically zero by Eq. 12, so the claimed contribution is not supported by the paper's own math. read the letter →

arxiv 2507.18106 v1 pith:QATANDWG submitted 2025-07-24 cs.CV cs.AI

JinYoung Kim , DaeUng Jo , Kimin Yun , Jeonghyo Song , Youngjoon Yoo This is my paper · ORCID

classification cs.CVcs.AI

keywords out-of-distributiondetectionsemanticsegmentationuncertaintyestimationBetaposteriornetworknormalizingflowsfreeenergyresidualpatternlearningevidentialdeep

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 proposes a Free-Energy Posterior Network for semantic segmentation: a frozen backbone is paired with a flow-based density estimator that outputs per-pixel parameters of a Beta distribution, and the variance of that Beta distribution is used as a training signal for out-of-distribution (OoD) regions. The claim is that this variance map is a better supervisory target than the fixed free-energy scores used by the residual pattern learning (RPL) framework, because it lets the network learn to increase uncertainty near ambiguous or unseen regions instead of relying on post-hoc energy thresholding. If the claim holds, OoD-aware segmentation can be trained end-to-end in a single forward pass, with no stochastic sampling and no hand-tuned thresholds. The paper reports the best FPR and AuPRC across most of the Fishyscapes, RoadAnomaly, and Segment-Me-If-You-Can benchmarks.

What carries the argument

The operative object is the Beta posterior variance map. A flow-based network transforms per-pixel latent features into log-probability scores $z_{\mathrm{in}}, z_{\mathrm{out}}$, which are converted to Beta parameters by $\alpha = 1 + \log(1 + e^{z_{\mathrm{in}}})$ and $\beta = 1 + \log(1 + e^{z_{\mathrm{out}}})$; the variance $\mathrm{Var}(p) = \alpha\beta/((\alpha+\beta)^2(\alpha+\beta+1))$ then becomes both the uncertainty estimate and the training target. The key property is that variance is largest when $\alpha \approx \beta$ with small values, so minimizing the outlier loss corresponds to reducing the sharpness of the predictive distribution on OoD pixels. The paper integrates this variance into RPL's energy loss and adds a Beta Uncertainty Cross Entropy (BUCE) loss that combines standard cross-entropy, a digamma-based uncertainty term, and binary cross-entropy between the variance map and the OoD mask.

What would settle it

Run the same RPL-plus-flow pipeline but replace the Beta variance in the loss with the raw free-energy score while keeping the identical binary OoD-mask supervision; if FPR and AuPRC do not worsen, the Beta posterior variance is not the operative component. A second check is to recompute the flow log-likelihood with the full change-of-variables Jacobian determinant and see whether the free-energy ranking of OoD pixels in Eq. (5) is preserved.

Watch

Extended reading notes

Core claim

The paper's central claim is that distributional uncertainty, quantified as the variance of a per-pixel Beta posterior fitted by a normalizing-flow density estimator, can serve as a principled supervision signal for OoD detection in semantic segmentation. The method computes flow-based energy representations, converts them into Beta parameters $\alpha(\omega) = 1 + \log(1 + e^{z_{\mathrm{in}}(\omega)})$ and $\beta(\omega) = 1 + \log(1 + e^{z_{\mathrm{out}}(\omega)})$, and injects the variance $\mathrm{Var}(p(\omega)) = \frac{\alpha(\omega)\beta(\omega)}{(\alpha(\omega)+\beta(\omega))^2(\alpha(\omega)+\beta(\omega)+1)}$ into RPL's energy-based loss. Because Beta variance is maximized when $\alpha \approx \beta$ and both are small, the loss pushes the flow network to flatten the predictive distribution exactly where the OoD mask is active, while the proposed BUCE loss keeps inlier predictions confident. The claimed result is a semantically meaningful, sampling-free uncertainty map that outperforms MC Dropout, Shannon entropy, and plain free-energy baselines on real-world anomaly-segmentation benchmarks.

Load-bearing premise

The load-bearing assumption is that the variance of the Beta posterior, trained with a binary cross-entropy loss against OoD labels, truly captures distributional uncertainty for unseen pixels and is a better training target than the original fixed free-energy score; if the variance is merely fitting the outlier labels, the claimed mechanism collapses.

Editorial extensions

If this is right

OoD-aware segmentation can be trained with a single forward pass, eliminating the computational overhead of MC Dropout-style sampling and the need for post-hoc energy thresholding.
The Beta variance map provides dense pixel-level supervision, so the network can learn OoD regions during training rather than only at inference, which is what the reported FPR and AuPRC gains reflect.
The same uncertainty signal transfers across different anomaly types, since the paper reports consistent gains on Fishyscapes-Static, Fishyscapes-L&F, SMIYC-Anomaly, SMIYC-Obstacle, and RoadAnomaly.
Because the segmentation backbone is frozen and only the flow posterior and loss terms are trained, the added cost is localized to the uncertainty branch, making the approach more scalable than sampling-based alternatives.
The variance-maximizing condition ties OoD behavior to a concrete property of the Beta distribution, giving practitioners a diagnostic: anomalous pixels should be those where the posterior is flat and low-evidence.

Reading between the lines

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

The same Beta-variance supervision could be plugged into other energy-based OoD detectors beyond RPL; the paper's closing remark gestures at this, but the mechanism itself is not specific to RPL.
Part of the reported gain may come from the binary cross-entropy term that directly supervises variance with OoD labels, so an ablation that trains the same BCE on raw energy scores would reveal how much of the improvement is due to the Beta parametrization rather than to outlier exposure.
If Beta variance is a genuine distributional-uncertainty signal, it should also improve misclassification detection and calibration on in-distribution inputs; the paper does not report those checks, but they are testable extensions.
The condition $\alpha \approx \beta$ with small values is the evidential reading of 'conflicting and weak evidence,' which suggests the method could be unified with evidential deep-learning losses that penalize high uncertainty without explicit OoD labels.

Share X Bluesky LinkedIn Reddit HN

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

Here's the short version: the headline contribution is a no-op. In Sec. 3.1, Eq. 12 defines Lout = Σ max(−m(ω)·Var(p(ω)), 0). Since m is 0/1 and Var is always positive, every term is max(negative, 0) = 0. The gradient in Eq. 13 (∇ = −∂Var/∂α − ∂Var/∂β) is the derivative of −Var, not of the max expression, which is 0 on the whole domain. So the variance map injects no gradient into the flow network. Whatever improvement the experiments show has to come from the other loss terms (LUCE or Lvar), not from the mechanism the paper is built around.

What's actually new is narrow: a flow-based Beta posterior network, trained with a UCE-style loss plus a BCE variance term, plugged into RPL for semantic segmentation. The components are all prior work, but the assembly is new in a limited sense. The motivation — that RPL's fixed energy targets are too rigid — is reasonable, and the paper applies the idea to the right benchmarks.

The soft spots are severe. Eq. 5 is not the log-likelihood of a Gaussian under a normalizing flow; the log-determinant term is missing and the diagonal term should be log diag(U), with a normalization constant. So the free-energy interpretation is on shaky ground. Table 1 omits RPL, the method's own foundation, while the text claims to outperform all baselines. Table 3 shows RPL is actually better on Fishyscapes-Static FPR (0.85 vs 1.03) and on SMIYC-Obstacle (FPR 0.09→0.11, AUPRC 96.91→96.29). The text then misquotes numbers (e.g., 89.48 vs 57.99 and 96.29 vs 22.10) that don't match the table. There is no code and no hyperparameter values (λ1, λ2, τ, N), so the experiments can't be reproduced. Finally, the variance used in Eq. 18 is trained with BCE against the OoD labels, so calling it an unsupervised distributional uncertainty estimate is overstated.

Bottom line: this paper is not ready for a serious referee. The central loss is mathematically inert, and the experimental reporting is inconsistent. The authors could fix the loss, correct the flow equation, add the missing baseline, and release code — then there might be a modest, useful paper. As written, I would desk-reject and ask for a careful revision with those changes.

Referee Report

5 major / 5 minor

Summary. The paper proposes a Free-Energy Posterior Network that combines a Beta-posterior flow module with the RPL framework for out-of-distribution detection in semantic segmentation. It claims that using the variance of the Beta posterior as a supervisory signal, instead of RPL's fixed free-energy targets, improves uncertainty estimation and OoD detection. Experiments are reported on Fishyscapes, RoadAnomaly, and Segment-Me-If-You-Can.

Significance. The topic is relevant and the paper attempts a novel combination of posterior-network uncertainty and energy-based flow models. However, the central mechanism as written is non-functional: the proposed OoD variance loss in Eq. (12) is identically zero, and the displayed gradient in Eq. (13) is inconsistent with that loss. The empirical section also misreports a table and omits the most relevant baseline. As a result, the claimed contribution is not established by the evidence presented.

major comments (5)

[§3.1, Eqs. (11-13)] Eq. (12) defines Lout = sum_omega max(-m(omega)*Var(p(omega)), 0), with m(omega) in {0,1} and Var(p(omega)) > 0 for alpha,beta > 0. Every summand is therefore zero: for an OoD pixel the argument is negative and max(.,0) returns 0, while for an inlier pixel the argument is 0. Thus Lout is identically zero and injects no gradient into the flow parameters. The gradient in Eq. (13) is the derivative of -Var, not of max(-m*Var,0), whose gradient is zero on the whole domain. This invalidates the paper's central claim that the Beta variance map replaces RPL's fixed free-energy target as a supervisory signal; any observed difference between RPL and RPL+DE would have to be attributed to other terms, such as Eq. (18), not to Eq. (12).
[§3.2, Eq. (18); §4.2] The only operational supervision of the Beta variance is the binary cross-entropy loss Lvar = BCE(Var(i), y(i)), which directly fits the variance to binary OoD labels. The paper claims that the method estimates distributional uncertainty without relying on handcrafted thresholds, but Eq. (18) uses OoD labels as supervision. Consequently, reporting improved OoD detection does not establish that Beta variance captures intrinsic distributional uncertainty; it may simply reflect that the estimator is trained to predict the target labels. An ablation omitting Lvar or a held-out analysis would be needed to support the interpretation.
[§4.2, Table 3] The text in Section 4.2 misreports Table 3. For SMIYC-Anomaly, Table 3 shows RPL+DE with FPR 6.87 and AUPRC 89.48 versus RPL with FPR 7.18 and AUPRC 88.55, not AP 89.48 vs. 57.99. For SMIYC-Obstacle, Table 3 shows RPL+DE with FPR 0.11 and AUPRC 96.29 versus RPL with FPR 0.09 and AUPRC 96.91, not AUPRC 96.29 vs. 22.10. The numbers 57.99 and 22.10 do not appear in the table. Moreover, the Obstacle row shows that RPL+DE is worse than RPL on both FPR and AUPRC, contradicting the claim that RPL+DE consistently improves both AP and AUPRC.
[§4.1, Table 1; §4.2] Table 1 omits the RPL baseline even though the paper's contribution is an integration with RPL and Section 4.2 claims that the method outperforms all baselines. Without RPL numbers in Table 1, the comparison is incomplete, especially because Table 3 indicates that the gain over RPL is small or negative on some benchmarks. The claim should be restricted to the baselines actually listed in the table.
[§2.2, Eq. (5)] Eq. (5) is presented as the log-likelihood in a flow-based framework, but it appears to be only the log-density of the latent Gaussian. A normalizing-flow density for the input requires the log-absolute-determinant of the Jacobian of the transformation. If Eq. (5) is intended only as the latent density, the text should say so explicitly. Since the free-energy interpretation of the flow output is central to the proposed framework, this needs clarification.

minor comments (5)

[§3.1, Eqs. (7) and (10)] The definitions of alpha and beta are inconsistent: Eq. (7) uses log(1 + exp(z_in)) and log(1 + exp(z_out)), while Eq. (10) uses log P(z|c) and log Q(z|c). The relationship between these notations should be made explicit.
[§3.1, Eq. (9)] The threshold tau in Eq. (9) is not specified; if it is a hyperparameter, its value and sensitivity should be reported.
[§3.2, Eq. (15)] The variable y_i in Eq. (15) is used as a class indicator, but its exact meaning relative to Eq. (9) is unclear for the multi-class segmentation setting. Please clarify whether it is a pixel-wise binary OoD indicator or a class label.
[Figure 2 caption] The caption says that the differential entropy map replaces the free energy term in RPL's l_in loss, whereas Section 3.1 defines L_out for the OoD term. The notation should be reconciled.
[§4.3 and Figure 3] Section 4.3 refers to the Fishyscapes dataset, but the Figure 3 caption says Road Anomaly. Please correct the caption or the text.

Circularity Check

1 steps flagged · score 6.0 of 10

Eq. 12's variance loss is identically zero under the paper's own positivity of α,β, so the central variance-based supervision mechanism reduces to a constant; Eq. 13's gradient is inconsistent with Eq. 12.

self definitional [Section 3.1, Eqs. (11)–(13)]
"Lout = Σ_{ω∈Ω} max(−m(ω) · Var(p(ω)), 0), where m(ω) ∈ {0,1} is a binary mask indicating OoD pixels. ... ∇θflowLout = − ∂Var(p)/∂α · ∂α/∂θflow − ∂Var(p)/∂β · ∂β/∂θflow."

Since α=1+log(1+e^{z_in})>1 and β=1+log(1+e^{z_out})>1 by Eq. (7), Var(p)>0 for every pixel. Hence −m(ω)·Var(p(ω)) is non-positive for both m=0 and m=1, so each max(·,0) term is identically zero and Lout≡0 regardless of the flow parameters. The gradient in Eq. 13 is the derivative of −Var, not of max(−m·Var,0), which is zero on the whole domain. Thus the paper's central claim that this loss 'encourages the flow network to produce Beta parameters with higher variance for OoD regions' cannot operate: the variance-based supervision reduces by construction to a null signal. Any improvement must come from the separately supervised Lvar term, not from Eq. 12.

full rationale

The derivation chain for the paper's central claim—that Beta variance replaces RPL's fixed free-energy target as a supervision signal—breaks at Eq. 12. Because α=1+log(1+e^{z_in}) and β=1+log(1+e^{z_out}) are always >1, Var(p)>0, so max(−m·Var,0)=0 for both m=0 and m=1. Thus Lout is identically zero and the gradient shown in Eq. 13 is not the gradient of Eq. 12. This is a self-definitional reduction: the proposed loss term cannot produce the 'higher variance for OoD regions' behavior described in Sec. 3.1. The other uncertainty terms (LUCE, Lvar) are standard supervised losses trained on outlier-exposure labels and evaluated on held-out benchmarks, so they are not circular. There are no load-bearing self-citations: RPL and FlowEneDet are external prior works. The omitted Jacobian log-determinant in Eq. 5 is a correctness concern, not a circularity. Score reflects one central mechanism that reduces by construction to zero, while the empirical comparison remains externally benchmarked.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; it proposes a new loss and a new way to parameterize an existing Beta posterior, so the ledger is dominated by free hyperparameters and implicit modeling assumptions.

free parameters (5)

lambda_1
Weight of the UCE loss term in Eq (19); value never reported.
lambda_2
Weight of the variance consistency loss in Eq (19); value never reported.
tau
Threshold in Eq (9) for binarizing inlier/outlier predictions; not specified.
UCE scale factor = 1e-7
Ad hoc scaling of the UCE loss in Sec 3.2, chosen for numerical stability without justification.
N
Confidence scaling factor in Eq (2), referenced but never defined or set.

assumptions (3)

domain assumption Beta/Dirichlet posterior parameters encode distributional uncertainty.
Invoked in Sec 2.1 (Eq 1-4) without validating that the flow-based density estimates produce well-calibrated posterior parameters.
domain assumption Free energy computed from a normalizing flow is a valid OoD surrogate.
Sec 2.2 assumes low energy implies inlier, but Eq (5) is not a correct flow log-likelihood, so the energy interpretation is not established.
domain assumption The flow network trained on frozen backbone features captures the semantic structure needed for OoD detection.
The method relies on the frozen segmentation model's features being informative enough; no analysis is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distributional Uncertainty for Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/QATANDWG

@misc{pith2026250718106,
  author       = {Pith},
  title        = {Pith review of: Distributional Uncertainty for Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QATANDWG}},
  note         = {Machine review of arXiv:2507.18106}
}

read the original abstract

Estimating uncertainty from deep neural networks is a widely used approach for detecting out-of-distribution (OoD) samples, which typically exhibit high predictive uncertainty. However, conventional methods such as Monte Carlo (MC) Dropout often focus solely on either model or data uncertainty, failing to align with the semantic objective of OoD detection. To address this, we propose the Free-Energy Posterior Network, a novel framework that jointly models distributional uncertainty and identifying OoD and misclassified regions using free energy. Our method introduces two key contributions: (1) a free-energy-based density estimator parameterized by a Beta distribution, which enables fine-grained uncertainty estimation near ambiguous or unseen regions; and (2) a loss integrated within a posterior network, allowing direct uncertainty estimation from learned parameters without requiring stochastic sampling. By integrating our approach with the residual prediction branch (RPL) framework, the proposed method goes beyond post-hoc energy thresholding and enables the network to learn OoD regions by leveraging the variance of the Beta distribution, resulting in a semantically meaningful and computationally efficient solution for uncertainty-aware segmentation. We validate the effectiveness of our method on challenging real-world benchmarks, including Fishyscapes, RoadAnomaly, and Segment-Me-If-You-Can.

Figures

Figures reproduced from arXiv: 2507.18106 by the authors.

**Figure 2.** Overview of the proposed architecture. Outlier Exposure (OE) samples are generated by synthesizing inlier (Cityscapes) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Qualitative comparison of uncertainty maps on the Road Anomaly dataset. (a) Input image, (b) baseline RPL [13], (c) our proposed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[1]

Ancha, P

S. Ancha, P. R. Osteen, and N. Roy. Deep evidential un- certainty estimation for semantic segmentation under out-of- distribution obstacles. In 2024 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 6943–6951. IEEE, 2024

work page 2024
[2]

The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation

H. Blum, P.-E. Sarlin, J. Nieto, R. Siegwart, and C. Ca- dena. The fishyscapes benchmark: Measuring blind spots in semantic segmentation. arXiv preprint arXiv:1904.03215, 2019

work page Pith review arXiv 1904
[3]

R. Chan, K. Lis, S. Uhlemeyer, H. Blum, S. Honari, R. Sieg- wart, P. Fua, M. Salzmann, and M. Rottmann. Segment- meifyoucan: A benchmark for anomaly segmentation. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

work page 2021
[4]

Charpentier, D

B. Charpentier, D. Z ¨ugner, and S. G ¨unnemann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural informa- tion processing systems, 33:1356–1367, 2020

work page 2020
[5]

L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam. Encoder-decoder with atrous separable convolution for se- mantic image segmentation. In Proceedings of the Euro- pean conference on computer vision (ECCV) , pages 801– 818, 2018

work page 2018
[6]

Cordts, M

M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016

2016
[7]

Gal and Z

Y . Gal and Z. Ghahramani. Dropout as a bayesian approx- imation: Representing model uncertainty in deep learning. ICML, 2016

work page 2016
[8]

Gudovskiy, T

D. Gudovskiy, T. Okuno, and Y . Nakata. Concurrent misclas- sification and out-of-distribution detection for semantic seg- mentation via energy-based normalizing flow. InUncertainty in Artificial Intelligence, pages 745–755. PMLR, 2023

work page 2023

Show all 17 references

[9]

Kendall and Y

A. Kendall and Y . Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017

2017
[10]

Landgraf, M

S. Landgraf, M. Hillemann, K. Wursthorn, and M. Ulrich. U- ce: Uncertainty-aware cross-entropy for semantic segmenta- tion. arXiv preprint arXiv:2307.09947, 2023

2023 arXiv
[11]

T. Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ra- manan, P. Dollar, and C. L. Zitnick. Microsoft COCO: Com- mon objects in context. In ECCV, 2014

2014
[12]

K. Lis, K. Nakka, P. Fua, and M. Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 2152–2161, 2019

2019
[13]

Y . Liu, C. Ding, Y . Tian, G. Pang, V . Belagiannis, I. Reid, and G. Carneiro. Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1151–1161, 2023

2023
[14]

Malinin and M

A. Malinin and M. Gales. Predictive uncertainty estimation via prior networks. Advances in neural information process- ing systems, 31, 2018

2018
[15]

Mukhoti and Y

J. Mukhoti and Y . Gal. Evaluating bayesian deep learn- ing methods for semantic segmentation. arXiv preprint arXiv:1811.12709, 2018

2018 arXiv
[16]

Sensoy, L

M. Sensoy, L. Kaplan, and M. Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems, 31, 2018

2018
[17]

W. Zhao, J. Li, X. Dong, Y . Xiang, and Y . Guo. Seg- ment every out-of-distribution object. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3910–3920, 2024

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Ancha, P

S. Ancha, P. R. Osteen, and N. Roy. Deep evidential un- certainty estimation for semantic segmentation under out-of- distribution obstacles. In 2024 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 6943–6951. IEEE, 2024

work page 2024

[2] [2]

The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation

H. Blum, P.-E. Sarlin, J. Nieto, R. Siegwart, and C. Ca- dena. The fishyscapes benchmark: Measuring blind spots in semantic segmentation. arXiv preprint arXiv:1904.03215, 2019

work page Pith review arXiv 1904

[3] [3]

R. Chan, K. Lis, S. Uhlemeyer, H. Blum, S. Honari, R. Sieg- wart, P. Fua, M. Salzmann, and M. Rottmann. Segment- meifyoucan: A benchmark for anomaly segmentation. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021

work page 2021

[4] [4]

Charpentier, D

B. Charpentier, D. Z ¨ugner, and S. G ¨unnemann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural informa- tion processing systems, 33:1356–1367, 2020

work page 2020

[5] [5]

L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam. Encoder-decoder with atrous separable convolution for se- mantic image segmentation. In Proceedings of the Euro- pean conference on computer vision (ECCV) , pages 801– 818, 2018

work page 2018

[6] [6]

Cordts, M

M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016

2016

[7] [7]

Gal and Z

Y . Gal and Z. Ghahramani. Dropout as a bayesian approx- imation: Representing model uncertainty in deep learning. ICML, 2016

work page 2016

[8] [8]

Gudovskiy, T

D. Gudovskiy, T. Okuno, and Y . Nakata. Concurrent misclas- sification and out-of-distribution detection for semantic seg- mentation via energy-based normalizing flow. InUncertainty in Artificial Intelligence, pages 745–755. PMLR, 2023

work page 2023