Pith. sign in

REVIEW 3 major objections 5 minor 58 references

Uncertainty-Aware Likelihood Ratio Estimation for Pixel-Wise Out-of-Distribution Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An uncertainty-aware likelihood ratio estimator, built on an evidential classifier, reports the lowest average false positive rate (2.5%) among state-of-the-art methods for pixel-wise out-of-distribution detection on five driving…

desk verdict A clean, plausible method with correct math, but the SOTA performance claim is undermined by inconsistent per-method averaging in Table 1. read the letter →

arxiv 2508.00587 v1 pith:JFUBLDHO submitted 2025-08-01 cs.CV

Marc Hölle , Walter Kellermann , Vasileios Belagiannis This is my paper
classification cs.CV
keywords out-of-distributiondetectionpixel-wiseanomalysegmentationlikelihoodratioestimationevidentialdeeplearningoutlierexposureuncertaintysemanticautonomousdriving
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

This paper argues that pixel-wise out-of-distribution (OoD) detection in driving scenes suffers because rare in-distribution classes and truly unknown objects produce similar kinds of uncertainty. It proposes to train a binary classifier to distinguish in-distribution features from synthetic proxy OoD features, but to give that classifier an evidential (Dirichlet) output so it can express how much evidence supports each decision. The paper shows that this uncertainty-aware likelihood ratio estimator avoids the overconfident extrapolation of a standard binary-cross-entropy classifier in low-density regions. Evaluated on five standard benchmarks, it reports the lowest average false positive rate (2.5%) at 95% true positive rate while holding 90.91% average precision, with negligible overhead on a frozen backbone. If true, the practical consequence is a cheap, attachable detector that mistrusts rare but known classes less and flags truly unknown objects more reliably.

What carries the argument

The load-bearing identity is the likelihood ratio trick, $p_{\mathrm{out}}(x)/p_{\mathrm{in}}(x)=P(O=1\mid x)/P(O=0\mid x)$ under a uniform prior on the binary source variable $O$, which lets a neural classifier stand in for the intractable density ratio. The classifier is made evidential by placing a Dirichlet distribution $\mathrm{Dir}(p_i\mid\alpha_i)$ over the binary class probabilities, with $\alpha_i=e_i+1$ where $e_i$ are the exponentiated logits; the expected class probability is $\alpha_{i1}/S_i$ and the vacuity $\nu_i=2/S_i$ quantifies epistemic uncertainty. Training uses the evidential log loss plus a KL regularizer with an annealing coefficient. At inference the likelihood ratio score simplifies to $\alpha_{i1}/\alpha_{i0}$, which is the ratio of expected OoD evidence to expected ID evidence. This object carries the argument because it is what makes the model reduce confidence in low-density regions rather than extrapolate confidently.

What would settle it

Run an ablation where the BCE likelihood-ratio estimator is replaced by temperature-scaled BCE probabilities using the same $\alpha_{i1}/\alpha_{i0}$-style score; if its false positive rate matches the evidential estimator on datasets with high proxy-to-real overlap, the reported gains do not come from evidential uncertainty calibration. Alternatively, measure EDL vacuity against true distance to the training manifold in DINOv2 feature space and check whether the monotonic relationship the toy example predicts actually appears.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the likelihood ratio between in-distribution and out-of-distribution pixel features can be approximated by a binary classifier, and that training that classifier with an evidential loss instead of binary cross-entropy changes how it behaves outside the training support: instead of confidently extrapolating, it lowers its evidence and raises its vacuity. This yields a scoring function $\widehat{LR}(x_i)=\alpha_{i1}/\alpha_{i0}$ that consistently outperforms the BCE-trained likelihood ratio estimator across all five benchmarks, and the paper attributes the gain to the disentangling of aleatoric ambiguity from epistemic lack of evidence. The paper's headline result is an average false positive rate of 2.5% at 95% true positive rate, the lowest among compared methods, with 90.91% average precision.

Load-bearing premise

The central premise is that the uncertainty-calibration behaviour demonstrated on one-dimensional Gaussians transfers to the 1024-dimensional DINOv2 feature space, where the paper's only supporting evidence is a cosine-distance proxy to class means rather than a calibrated measure of true uncertainty.

Editorial extensions

If this is right

  • On the five benchmark datasets, the method reports the lowest average false positive rate (2.5%) at 95% true positive rate, with 90.91% average precision, beating the BCE-trained likelihood ratio estimator on every dataset.
  • The largest gains appear on Lost and Found No Known, where real obstacles differ most from the COCO-based synthetic outliers, supporting the claim that uncertainty awareness mitigates proxy bias.
  • The approach keeps the segmentation backbone frozen and adds negligible computational overhead, so it can be layered onto existing dense predictors without retraining them.
  • Because the estimator learns the ratio directly rather than modelling the two densities separately, it avoids the density-estimation failures that plague high-dimensional feature-space OoD detection.

Reading between the lines

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

  • If the calibration transfer holds, the same evidential likelihood-ratio scheme should apply to other dense prediction tasks with long-tailed classes, such as monocular depth or panoptic segmentation, where background and tail classes create similar false alarms.
  • The paper's extrapolation analysis uses cosine distance to the nearest class mean, which is not a calibrated uncertainty measure; a follow-up with a proper high-dimensional density or a held-out near-OoD dataset would sharpen the claim that vacuity tracks feature-space novelty.
  • A useful stress test would be to vary the synthetic outlier distribution (e.g., use ADE20K or a different domain gap) and check whether the EDL advantage grows precisely when the proxy distribution is a worse approximation to the true OoD distribution.
  • The method's score is a simple ratio of two Dirichlet concentrations, which might also permit uncertainty-aware calibration of the decision threshold, something the paper does not explore.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes an uncertainty-aware likelihood ratio estimator for pixel-wise out-of-distribution (OoD) detection in semantic segmentation. The method extracts dense DINOv2 features from a frozen segmentation model, trains a binary evidential classifier on Cityscapes in-distribution features and COCO-based synthetic outliers, and uses the ratio of expected Dirichlet class probabilities as an OoD score. The authors evaluate on five benchmarks (SMIYC Anomaly Track, Obstacle Track, Lost and Found NoKnown, Road Anomaly, Fishyscapes Static) and claim the lowest average false positive rate (2.5%) at 95% TPR among state-of-the-art methods, while outperforming a standard BCE likelihood ratio baseline.

Significance. If the claims hold, the method is practically attractive: it is a lightweight post-hoc detector that works with a frozen segmentation backbone, requires no retraining of the segmentation model, and adds negligible computational cost. The likelihood ratio derivation in Eqs. (1)-(4) is correct, the controlled univariate Gaussian experiment is a useful illustration, and the empirical comparison against the BCE baseline is reported on all five benchmarks. The paper also provides public code. However, the central SOTA claim rests on the average-FPR comparison in Table 1, and that comparison has a load-bearing flaw in how averages are computed; the claimed mechanism for the improvement also needs stronger support. The result is therefore significant but not yet conclusively established.

major comments (3)
  1. [Table 1 / Sec. 4.2.3] The 'Average' column in Table 1 is computed over different dataset subsets for different methods, so the averages are not directly comparable. For example, RbA's 3.87% FPR averages only three datasets, while EDL's 2.50% averages five. More concretely, the UNO row shows no entry for LaF NoKnown, but its reported average FPR of 3.52 cannot be reproduced from the four displayed entries: those four entries average to 2.40% FPR and 93.26% AP, not 3.52% and 94.24%. On the common four-dataset subset (Anomaly Track, Obstacle Track, Road Anomaly, FS Static), UNO's FPR of 2.40% is lower than EDL's 2.83% on that same subset. Thus the claim in the Abstract and Sec. 4.2.3 that EDL achieves the 'lowest average false positive rate' across all five benchmarks is not established. The authors should recompute all averages over a common intersection of datasets, report the number of datasets used for each average, and resolve the UNO discrepancy.
  2. [Sec. 4.2.2 / Eq. (14)] The comparison between the proposed EDL estimator and the BCE baseline varies both the training loss and the score function. For the BCE baseline, the likelihood ratio score is p/(1-p)=exp(o1-o0), while for the EDL estimator, Eq. (14) gives (exp(o1)+1)/(exp(o0)+1). These differ even for identical logits, particularly when evidence is low. The consistent improvement of EDL over BCE in Table 1 could therefore be due to the smoother +1 score transformation rather than to the evidential uncertainty modeling. To support the 'uncertainty-aware' claim, the authors should add ablations that hold the score function fixed (e.g., EDL loss with a BCE-style score, and BCE loss with an EDL-style score) or otherwise show that the effect is attributable to the evidential training.
  3. [Sec. 4.3.2] The extrapolation analysis intended to show that EDL reduces overconfident predictions in high-dimensional feature space uses cosine distance to the nearest class mean as a proxy for feature novelty. This is not a calibrated measure of epistemic uncertainty or of distance to the training distribution, and it does not directly validate that the univariate Gaussian behavior transfers to the 1024-dimensional DINOv2 feature space. The conclusion that EDL 'appropriately reduces predictive probability' in low-density regions should be supported by a more direct evaluation, such as reliability diagrams for the predicted probabilities, or a comparison of EDL and BCE on held-out ID features versus OoD features with a proper density or uncertainty estimate.
minor comments (5)
  1. [Sec. 4.2.2] The abstract and Sec. 5 claim 'negligible computational overhead', but no runtime, parameter count, or memory measurements are provided. Please add a quantitative complexity comparison.
  2. [Table 1] The table should state explicitly how the 'Average' column is computed and how many datasets contribute to each average; otherwise readers cannot judge the comparability of the values.
  3. [Sec. 4.3.1] The DINOv2 layer ablation in Fig. 3 is reported only on the SMIYC Anomaly Track validation set. Since the layer choice is a free parameter, please report the ablation across all five benchmarks or justify why a single benchmark is sufficient.
  4. [Fig. 2] The bottom panels of Fig. 2 are labeled 'Normalised Count', but it is not clear whether this refers to the density of training samples or to a histogram of predictions. Please clarify the definition in the caption.
  5. [Table 1] Minor formatting inconsistency: the BCE baseline's FPR on Road Anomaly is listed as '5.9' rather than '5.90', unlike the other entries. Please make the decimal places uniform.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the EDL likelihood-ratio estimator is trained on Cityscapes/COCO and evaluated on external benchmarks; the reported FPR/AP values are independent empirical results, not fitted constants.

full rationale

The paper's central derivation chain is self-contained: Eq. (2) is the standard likelihood-ratio trick, Eq. (4) turns a binary classifier's posterior into an odds ratio, and Eq. (14) evaluates that odds ratio under the evidential Dirichlet parameterization (alpha_1 / alpha_0). None of these steps defines the target result in terms of itself; the EDL score is a mathematically derived scoring function, not a fitted parameter renamed as a prediction. The estimator is trained on Cityscapes in-distribution features and COCO-based synthetic outliers, then evaluated on five external benchmarks (SMIYC Anomaly/Obstacle, LaF NoKnown, Road Anomaly, Fishyscapes Static) with real unknown objects and compared against prior published numbers and a same-architecture BCE baseline. The claimed advantage over the BCE baseline is supported by both a controlled univariate-Gaussian experiment and the benchmark table; it is not guaranteed by construction. Self-citations appear only in related work (e.g., references [11], [23-25], [34] with overlapping authors) and are not load-bearing for the main claim. The one significant caveat is not circularity: the 'lowest average FPR' headline rests on Table 1 averages computed over per-method dataset subsets, and the UNO row's average (AP 94.24, FPR 3.52) is arithmetically inconsistent with its four displayed entries (which average to approximately AP 93.26 and FPR 2.40). This affects the fairness of the SOTA ranking and should be corrected by recomputing averages over a common dataset intersection, but it is an aggregation/statistical-comparison issue, not a case of the prediction being equivalent to its inputs. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard statistical facts, the domain assumption that synthetic outliers transfer to real unknown objects, and the assumption that EDL's calibration behavior on a 1D example transfers to high-dimensional DINOv2 features. No new entities are postulated. The layer-21 choice and post-processing sigma are tuned values that affect every reported number.

free parameters (4)
  • DINOv2 feature layer index = 21
    Selected via ablation on the SMIYC Anomaly Track validation set (Fig 3); all reported results use layer 21, so this is a tuning choice on the evaluation benchmark.
  • Gaussian blur sigma for OoD score = 1
    Applied to upscaled scores before evaluation (Sec 4.2.2); affects FPR and AP and is not swept.
  • EDL KL annealing coefficient lambda_t = min(1, t/10)
    Schedule for the evidence regularization term in Eq 13; hand-set following prior EDL practice.
  • Adam learning rate = 2e-5
    Optimizer setting for the estimator, 10 epochs (Sec 4.2.2); no sensitivity analysis reported.
assumptions (6)
  • standard math Neyman-Pearson lemma: the likelihood ratio test is the uniformly most powerful test for a binary hypothesis test.
    Invoked in Sec 3.1 to justify using LR(xi) as the OoD score; the lemma concerns true densities, while the paper uses a proxy OoD distribution.
  • standard math Bayes' rule and a uniform prior P(O=0)=P(O=1) make the density ratio equal to posterior odds.
    Eq 2 in Sec 3.1; the paper notes a non-uniform prior is absorbed into the decision threshold.
  • domain assumption Synthetic COCO objects pasted into Cityscapes images are a useful proxy for true out-of-distribution features.
    Sec 3.2: the true OoD distribution is unknown, so AnomalyMix provides proxy features; the entire method depends on this transfer.
  • domain assumption Evidential deep learning's Dirichlet parameterization (alpha = exp(o)+1) and vacuity capture the uncertainty that matters for likelihood ratio estimation.
    Sec 3.4: the EDL machinery is imported from Sensoy et al. and used to train the binary classifier; no derivation shows it is optimal for correcting proxy bias.
  • domain assumption Frozen DINOv2 layer-21 features are sufficiently discriminative between known and unknown pixels.
    Sec 4.2.2 and Fig 3: backbone and layer choice are taken from prior work and an ablation, respectively.
  • domain assumption Cosine distance to the nearest class mean is a valid proxy for feature novelty in 1024 dimensions.
    Sec 4.3.2: used to support the extrapolation analysis; this is an auxiliary analysis, not the main benchmark evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Likelihood Ratio Estimation for Pixel-Wise Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/JFUBLDHO

@misc{pith2026250800587,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Likelihood Ratio Estimation for Pixel-Wise Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFUBLDHO}},
  note         = {Machine review of arXiv:2508.00587}
}
read the original abstract

Semantic segmentation models trained on known object classes often fail in real-world autonomous driving scenarios by confidently misclassifying unknown objects. While pixel-wise out-of-distribution detection can identify unknown objects, existing methods struggle in complex scenes where rare object classes are often confused with truly unknown objects. We introduce an uncertainty-aware likelihood ratio estimation method that addresses these limitations. Our approach uses an evidential classifier within a likelihood ratio test to distinguish between known and unknown pixel features from a semantic segmentation model, while explicitly accounting for uncertainty. Instead of producing point estimates, our method outputs probability distributions that capture uncertainty from both rare training examples and imperfect synthetic outliers. We show that by incorporating uncertainty in this way, outlier exposure can be leveraged more effectively. Evaluated on five standard benchmark datasets, our method achieves the lowest average false positive rate (2.5%) among state-of-the-art while maintaining high average precision (90.91%) and incurring only negligible computational overhead. Code is available at https://github.com/glasbruch/ULRE.

Figures

Figures reproduced from arXiv: 2508.00587 by the authors.

Figure 1
Figure 1. Visualisations for pixel-wise likelihood ratio. Overlay of predicted pixel-wise likelihood ratios on images with unknown objects from diverse driving scenes. Our method successfully detects unknown objects without false detections on in-distribution regions. Images are taken from the Segment-Me-If-You-Can [7] and Lost and Found no Known [41] datasets. Abstract Semantic segmentation models trained on known object cla… view at source ↗
Figure 2
Figure 2. Visualisation of univariate Gaussian classification. Top: Comparison between an evidential classifier (EDL) and a standard binary classifier trained with binary cross-entropy (BCE), with corresponding uncertainty measures, vacuity and entropy, respectively. Shaded blue area indicates the training data den￾sity. Bottom: Predicted Dirichlet distributions Dir(p|α) at x ∈ {−6, −0.4, 0, 0.4, 6}. The vertical dashed lines… view at source ↗
Figure 3
Figure 3. DINOv2 layer ablation study. Comparison of models trained using features from different layers l ∈ {1, 6, 12, 16, 21, 24} of DINOv2. Performance is reported in terms of average precision (↑) and false positive rate (↓) on the SMIYC Anomaly Track validation set. 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Minimum Distance mink dcos(µk ,xi) 0.0 0.2 0.4 0.6 0.8 1.0 Average Probability ¯p θ (yout i = 1|xi) p¯BCE θ (y out i = 1|xi) p¯ED… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualisation of overconfident extrapolation. Com￾parison between an evidential classifier (EDL) and a standard bi￾nary classifier (BCE) on the Fishyscapes Static dataset. The plot shows the average probability p¯θ(y out = 1|xi) over the cosine distance dcos(µk , xi) b…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 49 canonical work pages

  1. [1]

    Maskomaly:Zero-Shot Mask Anomaly Segmentation

    Jan Ackermann, Christos Sakaridis, and Fisher Yu. Masko- maly: Zero-shot mask anomaly segmentation. arXiv preprint arXiv:2305.16972, 2023. 8

    work page Pith review arXiv 2023

  2. [2]

    Deep evidential uncertainty estimation for semantic segmentation under out-of-distribution obstacles

    Siddharth Ancha, Philip R Osteen, and Nicholas Roy. Deep evidential uncertainty estimation for semantic segmentation under out-of-distribution obstacles. In 2024 IEEE Inter- national Conference on Robotics and Automation (ICRA) , pages 6943–6951. IEEE, 2024. 3

    work page 2024

  3. [3]

    Towards open set deep networks

    Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1563–1572,

    work page

  4. [4]

    Pattern recognition and machine learning

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 4

    work page 2006

  5. [5]

    Breaking down out-of-distribution de- tection: Many methods based on ood training data estimate a combination of the same core quantities

    Julian Bitterwolf, Alexander Meinke, Maximilian Augustin, and Matthias Hein. Breaking down out-of-distribution de- tection: Many methods based on ood training data estimate a combination of the same core quantities. In International conference on machine learning, pages 2041–2074. PMLR,

    work page

  6. [6]

    The fishyscapes benchmark: Measuring blind spots in semantic segmentation

    Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. The fishyscapes benchmark: Measuring blind spots in semantic segmentation. Inter- national Journal of Computer Vision , 129(11):3119–3135,

    work page

  7. [7]

    Segmentmeifyou- can: A benchmark for anomaly segmentation

    Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Math- ieu Salzmann, and Matthias Rottmann. Segmentmeifyou- can: A benchmark for anomaly segmentation. arXiv preprint arXiv:2104.14812, 2021. 1, 2, 6, 8

    arXiv 2021

  8. [8]

    Entropy maximization and meta classification for out-of- distribution detection in semantic segmentation

    Robin Chan, Matthias Rottmann, and Hanno Gottschalk. Entropy maximization and meta classification for out-of- distribution detection in semantic segmentation. In Proceed- ings of the ieee/cvf international conference on computer vi- sion, pages 5128–5137, 2021. 2, 8

    work page 2021

Show all 58 references
  1. [9]

    Per- pixel classification is not all you need for semantic segmen- tation

    Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per- pixel classification is not all you need for semantic segmen- tation. Advances in neural information processing systems , 34:17864–17875, 2021. 3

    2021

  2. [10]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceed- ings of the IEEE conference on computer vision and pattern re...

    2016

  3. [11]

    Diffusion model guided sampling with pixel-wise aleatoric uncertainty esti- mation

    Michele De Vita and Vasileios Belagiannis. Diffusion model guided sampling with pixel-wise aleatoric uncertainty esti- mation. In 2025 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV), pages 3844–3854. IEEE,

    2025

  4. [12]

    Outlier detection by ensembling uncertainty with negative objectness

    Anja Deli ´c, Matej Grci´c, and Siniˇsa ˇSegvi´c. Outlier detection by ensembling uncertainty with negative objectness. arXiv preprint arXiv:2402.15374, 2024. 1, 3, 7, 8

    2024 arXiv

  5. [13]

    Pixel-wise anomaly detection in complex driving scenes

    Giancarlo Di Biase, Hermann Blum, Roland Siegwart, and Cesar Cadena. Pixel-wise anomaly detection in complex driving scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16918– 16927, 2021. 2, 8

    2021

  6. [14]

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

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning , pages 1050–1059. PMLR, 2016. 3

    2016

  7. [15]

    Far away in the deep space: dense nearest-neighbor-based out-of- distribution detection

    Silvio Galesso, Max Argus, and Thomas Brox. Far away in the deep space: dense nearest-neighbor-based out-of- distribution detection. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 4477– 4487, 2023. 2, 8

    2023

  8. [16]

    Diffusion for out-of-distribution detection on road scenes and beyond

    Silvio Galesso, Philipp Schr ¨oppel, Hssan Driss, and Thomas Brox. Diffusion for out-of-distribution detection on road scenes and beyond. In European Conference on Computer Vision, pages 110–126. Springer, 2024. 2

    2024

  9. [17]

    Densehy- brid: Hybrid anomaly detection for dense open-set recogni- tion

    Matej Grci ´c, Petra Bevandi ´c, and Sini ˇsa ˇSegvi´c. Densehy- brid: Hybrid anomaly detection for dense open-set recogni- tion. In European Conference on Computer Vision , pages 500–517. Springer, 2022. 2, 8

    2022

  10. [18]

    On advantages of mask-level recognition for outlier-aware segmentation

    Matej Grci ´c, Josip ˇSari´c, and Siniˇsa ˇSegvi´c. On advantages of mask-level recognition for outlier-aware segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2937–2947, 2023. 3, 7, 8

    2023

  11. [19]

    Dense out-of-distribution detection by robust learn- ing on synthetic negative data

    Matej Grci ´c, Petra Bevandi ´c, Zoran Kalafati ´c, and Sini ˇsa ˇSegvi´c. Dense out-of-distribution detection by robust learn- ing on synthetic negative data. Sensors, 24(4):1248, 2024. 2, 8

    2024

  12. [20]

    Identifying out-of-domain objects with dirichlet deep neural networks

    Ahmed Hammam, Frank Bonarens, Seyed Eghbal Ghobadi, and Christoph Stiller. Identifying out-of-domain objects with dirichlet deep neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4560–4569, 2023. 3

    2023

  13. [21]

    Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the prob- lem

    Matthias Hein, Maksym Andriushchenko, and Julian Bitter- wolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the prob- lem. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4...

    2019

  14. [22]

    Deep anomaly detection with outlier exposure

    Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018. 1, 2, 4

    2018 arXiv

  15. [23]

    Heatmap- based out-of-distribution detection

    Julia Hornauer and Vasileios Belagiannis. Heatmap- based out-of-distribution detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2603–2612, 2023. 3

    2023

  16. [24]

    Out-of-distribution detection for monocular depth esti- mation

    Julia Hornauer, Adrian Holzbock, and Vasileios Belagian- nis. Out-of-distribution detection for monocular depth esti- mation. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 1911–1921, 2023. 2

    1911

  17. [25]

    Revisiting gradient-based uncertainty for monocular depth estimation

    Julia Hornauer, Amir El-Ghoussani, and Vasileios Belagian- nis. Revisiting gradient-based uncertainty for monocular depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3

    2025

  18. [26]

    Subjective Logic: A formalism for reasoning under uncertainty

    Audun Jsang. Subjective Logic: A formalism for reasoning under uncertainty. Springer Publishing Company, Incorpo- rated, 2018. 4 9

    2018

  19. [27]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  20. [28]

    DEUP: Direct epistemic uncertainty pre- diction

    Salem Lahlou, Moksh Jain, Hadi Nekoei, Victor I Butoi, Paul Bertin, Jarrid Rector-Brooks, Maksym Korablyov, and Yoshua Bengio. DEUP: Direct epistemic uncertainty pre- diction. Transactions on Machine Learning Research, 2023. Expert Certification. 3

    2023

  21. [29]

    Simple and scalable predictive uncertainty estima- tion using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estima- tion using deep ensembles. Advances in neural information processing systems, 30, 2017. 3

    2017

  22. [30]

    A simple unified framework for detecting out-of-distribution samples and adversarial attacks

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural infor- mation processing systems, 31, 2018. 2

    2018

  23. [31]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceeding...

    2014

  24. [32]

    Detecting the unexpected via image resynthesis

    Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152–2161, 2019. 2, 6, 8

    2019

  25. [33]

    Detecting road obstacles by erasing them.IEEE trans- actions on pattern analysis and machine intelligence, 46(4): 2450–2460, 2023

    Krzysztof Lis, Sina Honari, Pascal Fua, and Mathieu Salz- mann. Detecting road obstacles by erasing them.IEEE trans- actions on pattern analysis and machine intelligence, 46(4): 2450–2460, 2023. 8

    2023

  26. [34]

    Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation

    Yuyuan Liu, Choubo Ding, Yu Tian, Guansong Pang, Vasileios Belagiannis, Ian Reid, and Gustavo 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 ,...

    2023

  27. [35]

    Rba: Segmenting unknown regions rejected by all

    Nazir Nayal, Misra Yavuz, Joao F Henriques, and Fatma G¨uney. Rba: Segmenting unknown regions rejected by all. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 711–722, 2023. 2, 3, 4, 7, 8

    2023

  28. [36]

    A likelihood ratio-based approach to segmenting unknown objects

    Nazir Nayal, Youssef Shoeb, and Fatma G¨uney. A likelihood ratio-based approach to segmenting unknown objects. arXiv preprint arXiv:2409.06424, 2024. 1, 2, 3, 6, 8

    2024 arXiv

  29. [37]

    The mapillary vistas dataset for semantic understanding of street scenes

    Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision , pages 4990– 4999, 2017. 7

    2017

  30. [38]

    Jerzy Neyman and Egon Sharpe Pearson. Ix. on the problem of the most efficient tests of statistical hypotheses. Philo- sophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Char- acter, 231(694-706):289–337, 1933. 3

    1933

  31. [39]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 8

    2023 arXiv

  32. [40]

    Learn to accumulate evi- dence from all training samples: theory and practice

    Deep Shankar Pandey and Qi Yu. Learn to accumulate evi- dence from all training samples: theory and practice. In In- ternational Conference on Machine Learning, pages 26963– 26989. PMLR, 2023. 4, 5

    2023

  33. [41]

    Lost and found: detecting small road hazards for self-driving vehi- cles

    Peter Pinggera, Sebastian Ramos, Stefan Gehrig, Uwe Franke, Carsten Rother, and Rudolf Mester. Lost and found: detecting small road hazards for self-driving vehi- cles. In 2016 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pages 1099–1106. IEEE,

    2016

  34. [42]

    Ghost: Gaussian hypothesis open-set tech- nique

    Ryan Rabinowitz, Steve Cruz, Manuel G ¨unther, and Ter- rance E Boult. Ghost: Gaussian hypothesis open-set tech- nique. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6666–6674, 2025. 2

    2025

  35. [43]

    Unmasking anomalies in road- scene segmentation

    Shyam Nandan Rai, Fabio Cermelli, Dario Fontanel, Carlo Masone, and Barbara Caputo. Unmasking anomalies in road- scene segmentation. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4037–4046,

  36. [44]

    Likelihood ratios for out-of-distribution detec- tion

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshmi- narayanan. Likelihood ratios for out-of-distribution detec- tion. Advances in neural information processing systems , 32, 2019. 2

    2019

  37. [45]

    Learning likelihood ratios with neural network classifiers

    Shahzar Rizvi, Mariel Pettee, and Benjamin Nachman. Learning likelihood ratios with neural network classifiers. Journal of High Energy Physics, 2024(2):1–41, 2024. 3

    2024

  38. [46]

    Toward open set recogni- tion

    Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recogni- tion. IEEE transactions on pattern analysis and machine intelligence, 35(7):1757–1772, 2012. 1

    2012

  39. [47]

    Eviden- tial deep learning to quantify classification uncertainty

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Eviden- tial deep learning to quantify classification uncertainty. Ad- vances in neural information processing systems , 31, 2018. 3, 4, 5

    2018

  40. [48]

    Density ratio estimation in machine learning

    Masashi Sugiyama, Taiji Suzuki, and Takafumi Kanamori. Density ratio estimation in machine learning . Cambridge University Press, 2012. 2, 3

    2012

  41. [49]

    Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes

    Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuan- hong Chen, and Gustavo Carneiro. Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In European Conference on Computer Vision, pages 246–263. Springer, 2022. 1, 2, 3, 4, 6, 8

    2022

  42. [50]

    Prior and posterior networks: A survey on evidential deep learn- ing methods for uncertainty estimation

    Dennis Ulmer, Christian Hardmeier, and Jes Frellsen. Prior and posterior networks: A survey on evidential deep learn- ing methods for uncertainty estimation. Transactions on Ma- chine Learning Research, 2023. 3

    2023

  43. [51]

    Image-consistent detection of road anomalies as unpredictable patches

    Tom ´aˇs V oj´ıˇr and Ji ˇr´ı Matas. Image-consistent detection of road anomalies as unpredictable patches. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 5491–5500, 2023. 8

    2023

  44. [52]

    Pixood: Pixel- level out-of-distribution detection

    Tom ´aˇs V oj´ıˇr, Jan ˇSochman, and Ji ˇr´ı Matas. Pixood: Pixel- level out-of-distribution detection. In European Conference on Computer Vision, pages 93–109. Springer, 2024. 2, 3, 6, 7, 8

    2024

  45. [53]

    Falsehoods that ml re- searchers believe about ood detection

    Andi Zhang and Damon Wischik. Falsehoods that ml re- searchers believe about ood detection. InNeurIPS ML Safety Workshop, 2022. 2, 3 10

    2022

  46. [54]

    Csl: Class-agnostic structure-constrained learn- ing for segmentation including the unseen

    Hao Zhang, Fang Li, Lu Qi, Ming-Hsuan Yang, and Naren- dra Ahuja. Csl: Class-agnostic structure-constrained learn- ing for segmentation including the unseen. InProceedings of the AAAI Conference on Artificial Intelligence, pages 7078– 7086, 2024. 8

    2024

  47. [55]

    Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments

    Jingyang Zhang, Nathan Inkawhich, Randolph Linderman, Yiran Chen, and Hai Li. Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 5531–5540, 2023. 4

    2023

  48. [56]

    Under- standing failures in out-of-distribution detection with deep generative models

    Lily Zhang, Mark Goldstein, and Rajesh Ranganath. Under- standing failures in out-of-distribution detection with deep generative models. In International Conference on Machine Learning, pages 12427–12436. PMLR, 2021. 2

    2021

  49. [57]

    Segment every out-of-distribution object

    Wenjie Zhao, Jia Li, Xin Dong, Yu Xiang, and Yunhui Guo. Segment every out-of-distribution object. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3910–3920, 2024. 7

    2024

  50. [58]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 633–641,

Pith tools

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