Pith. sign in

REVIEW 4 major objections 4 minor 79 references

Anomalies by Synthesis: Anomaly Detection using Generative Diffusion Models for Off-Road Navigation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A diffusion model can reveal anomalies in off-road images by synthesizing an edited, anomaly-free version of the scene, then comparing the two.

desk verdict A useful off-road anomaly-detection pipeline whose core guidance trick is a known DPS-style heuristic; the paper overstates novelty and needs stronger experiments, but it deserves a serious referee. read the letter →

arxiv 2505.22805 v1 pith:QXPEICIB submitted 2025-05-28 cs.RO cs.CVcs.LG

Siddharth Ancha , Sunshine Jiang , Travis Manderson , Laura Brandt , Yilun Du , Philip R. Osteen , Nicholas Roy This is my paper
classification cs.ROcs.CVcs.LG
keywords anomalydetectiondiffusionmodelsanalysisbysynthesisguidedoff-roadnavigationout-of-distributionCLIPfeaturesvision-language
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 that a robot can find unknown obstacles in off-road images by generating an edited version of the scene and comparing the two. The generative model, a diffusion model trained only on normal terrain, is guided at test time to synthesize a new image that stays close to the input but is more probable under the training distribution, which removes out-of-distribution objects such as people, vehicles, and buildings. Anomalies are then located by comparing pixels of the input and edited images in a learned semantic feature space rather than by raw intensity, so subtle whole-image changes do not create false alarms. The paper claims this works with no assumptions about what anomalies look like, no out-of-distribution training data, and no retraining of the diffusion model.

What carries the argument

The load-bearing mechanism is a generalized similarity-conditioned guidance gradient for reverse diffusion. Instead of following classifier guidance, the paper tracks the ideal gradient $\nabla_{x_t} \log \mathbb{E}_{q(x_0 \mid x_t)}[r_{\mathrm{sim}}(x_0, x^{\mathrm{input}}_0)]$ and replaces the intractable expectation by evaluating the similarity at the diffusion model's own prediction $\mu^\theta_0(x_t)$ of the clean image, which is obtained from the noise-prediction network; this requires a backward pass through the network. The analysis side is carried by comparing the input and edited images in the CLIP feature space at full resolution, using MaskCLIP features upsampled by FeatUp and cleaned with SAM segments, so that semantically meaningful edits are scored while tiny RGB changes are ignored.

What would settle it

Run the same input image through reverse diffusion with two guidance gradients: the paper's $g^\theta_t(x_t)$ and a Monte Carlo estimate of $\nabla_{x_t} \log \mathbb{E}_{q(x_0 \mid x_t)}[r_{\mathrm{sim}}(x_0, x^{\mathrm{input}}_0)]$ obtained by sampling many $x_0$ from $q(x_0 \mid x_t)$, then compare the resulting edited images. If the gradients diverge substantially at intermediate timesteps where anomalies are present, the point-estimate approximation is not tracking the ideal edit and the theoretical grounding of the pipeline fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that pixel-wise anomaly detection can be reduced to analysis-by-synthesis: sampling from the conditional distribution $q(x^{\mathrm{edit}}_0 \mid x^{\mathrm{input}}_0) \propto q(x^{\mathrm{edit}}_0)\, r_{\mathrm{sim}}(x^{\mathrm{edit}}_0, x^{\mathrm{input}}_0)$, then reading off the edit. The theoretical contribution is Theorem 1, which states that the correct guidance gradient during reverse diffusion is $\nabla_{x_t} \log \mathbb{E}_{q(x_0 \mid x_t)}[r_{\mathrm{sim}}(x_0, x^{\mathrm{input}}_0)]$, and the practical contribution is the approximation $g^\theta_t(x_t) = \nabla_{x_t} \log r_{\mathrm{sim}}(\mu^\theta_0(x_t), x^{\mathrm{input}}_0)$, where $\mu^\theta_0(x_t)$ is the diffusion model's predicted clean image at step $t$. The paper argues this point-estimate approximation is principled because it compares the input image to a denoised estimate at the same zero-noise level, and because it vanishes at $t \approx T$, where the true gradient is zero. It then claims that combining this guided editing with CLIP-feature comparison and SAM-based segment refinement yields accurate detection of small and multiple anomalies on the RUGD and RELLIS off-road datasets.

Load-bearing premise

The whole editing argument rests on the assumption that the reverse-diffusion posterior $q(x_0 \mid x_t)$ is concentrated enough around its mean that point-estimating it by $\mu^\theta_0(x_t)$ preserves the guidance gradient; if that is false at intermediate noise levels, the synthesized edit is no longer sampling from the stated distribution.

Editorial extensions

If this is right

  • A diffusion model trained only on ordinary off-road imagery can be repurposed for anomaly detection at test time, so the approach inherits improvements in generative modeling without needing to retrain.
  • Detection is class-agnostic: anything the diffusion model edits away, whether a vehicle, person, fish, or camouflaged object, is flagged without prior knowledge of the anomaly class.
  • Because edits are localized by SAM segments and scored in a semantic feature space, small and multiple anomalies can be found in a single image.
  • The same pipeline shows qualitative success on underwater imagery and can be accelerated with DDIM sampling, making it plausible for onboard robot use.

Reading between the lines

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

  • If the guidance-gradient approximation is sound, it generalizes beyond anomaly detection to any test-time image-editing task that projects an input onto a learned data manifold under a similarity constraint, such as inpainting or safety filtering of camera feeds.
  • The comparison feature space is not fixed by the method; CLIP could be replaced by newer vision-language features, so the framework may improve as foundation models improve.
  • A testable consequence of the paper's logic is that anomaly recall should be highest when the diffusion model's learned manifold is sharp near in-distribution content; datasets with highly varied normal terrain should stress the method more than the relatively uniform splits used in the paper.
  • The reported runtime of about 18 seconds per image with DDPM means inference speed is the practical bottleneck, and whether the 10x DDIM speedup preserves small-anomaly recall is an open empirical question raised by the paper.
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

4 major / 4 minor

Summary. The paper proposes an analysis-by-synthesis method for pixel-wise anomaly detection in off-road images. A DDPM trained on in-distribution images is used at test time to synthesize an edited image that is similar to the input but has out-of-distribution segments removed; anomaly scores are then obtained by comparing the input and edited images in a CLIP/MaskCLIP feature space, with FeatUp upsampling and SAM segment refinement. The paper derives a generalized similarity-conditioned guidance gradient (Theorem 1), approximates the intractable expectation by a point estimate at the diffusion model's predicted clean image (Eq. 3), and validates the pipeline on RUGD and RELLIS with additional qualitative underwater results.

Significance. If the approximation in Eq. 3 were controlled and the empirical results were robust, the paper would offer a useful test-time framework for off-road anomaly detection that requires no OOD training data and can reuse pretrained diffusion and vision-language models. The proof of Theorem 1 is concise, the pipeline is modular, and the appendix plus project website provide proofs and code access. However, the central inference approximation is not validated, the claimed quantitative superiority is contradicted by the paper's own ablation table, and the novelty claim overlooks closely related diffusion-posterior-sampling work. These issues currently prevent the paper from supporting its main claims.

major comments (4)
  1. [Section V, Eq. (3)] The replacement of E_{q(x0|xt)}[rsim(x0,xinput)] by rsim(mu^theta_0(xt), xinput) is uncontrolled, so the claim that this is a 'principled approximation' is not established. Even in the tractable case q(x0|xt)=N(mu, sigma_t^2 I), the exact expectation is (1+2*lambda*sigma_t^2)^(-d/2) * exp(-lambda*||mu-xinput||^2 / (1+2*lambda*sigma_t^2)), so the exact guidance gradient carries an attenuation factor 1/(1+2*lambda*sigma_t^2) that Eq. (3) drops. At intermediate-to-high noise levels, where edits are decided, this can overstate the guidance by a large, state-dependent factor, and with non-isotropic posterior covariance the error is direction-dependent. The t approximately T argument in Section V only checks an endpoint and gives no bound or diagnostic for intermediate t; since every edit is driven by this gradient, the theoretical grounding of the method is not supported as stated.
  2. [Section VII.C, Table I] The statement that the method 'outperforms all baselines and ablations, on both datasets and evaluation metrics' is contradicted by the reported numbers. The main method with FeatUp (row 15) achieves RUGD F*1 = 0.599 and RUGD AUC-PR = 0.709, while the ablation without FeatUp (row 14) achieves RUGD F*1 = 0.858 and RUGD AUC-PR = 0.724; row 14 is also better than row 15 on RELLIS F*1 (0.549 vs 0.540). The text mentions only that FeatUp may be useful for high-resolution visualizations; it does not reconcile this with the headline segmentation metric or explain why the main configuration is preferred when it is worse on that metric.
  3. [Section VII, Table I] No error bars, number of seeds, or significance tests are reported, and the guidance-strength hyperparameter selection protocol is not given. On RUGD AUC-PR the best reported number (0.724 without FeatUp) is only 0.019 above the nearest-neighbor baseline (0.705), and on RELLIS AUC-PR the main method's advantage over the normalizing-flow baseline (0.568 vs 0.541) is small. Because diffusion sampling is stochastic and the SAM/FeatUp components add variability, the reported single-run differences cannot be assessed as robust improvements; the authors should provide repeated-run statistics and a clear validation protocol for lambda/alpha.
  4. [Section V, novelty claim] The sentence 'As far as we are aware, ours is the first work to leverage the learned diffusion model epsilon^theta_t(xt) to compute the guidance gradient' omits Diffusion Posterior Sampling (DPS; Chung et al., NeurIPS 2022), which uses the same score-substitution idea of backpropagating through the predicted clean image xhat_0 obtained from the learned diffusion model. Even if the application domain and conditioning function differ, the inference technique is not new; the paper should cite DPS and clearly differentiate its contribution from that line of work.
minor comments (4)
  1. [Appendix, Fig. 4 caption] The caption says 'q(xt | xt) = N(xt; sqrt(1-beta_t) xt-1, beta_t I)', but the left-hand side should read q(xt | xt-1); as written, the conditioning variable does not match the formula.
  2. [Section V] The statement that 'in forward diffusion, x0 and xT are independently distributed' is only true asymptotically as T grows under the assumed noise schedule; for finite T the joint q(x0, xT) does not factor exactly. Please rephrase as an approximation.
  3. [Section VII.C] The guidance strength is called lambda in Eq. (1) and Algorithm 1 but alpha in Appendix II.C and Fig. 8; the relationship between these notations and the actual tuning procedure should be stated explicitly.
  4. [Section VII.C] The paper should state clearly whether FeatUp is part of the recommended method or only for visualization, since Table I shows that removing FeatUp improves several quantitative metrics; the current text leaves this ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the guidance target distribution is externally defined, and anomaly evaluation uses held-out ground-truth labels.

full rationale

The paper's derivation chain is self-contained and externally anchored. The synthesis objective in Eq. (1) is defined before any anomaly labels are used: q(xedit|xinput) is proportional to q(xedit), approximated by a diffusion model trained only on in-distribution images, times a fixed Gaussian similarity kernel rsim. Theorem 1 is proven in Appendix I from the score decomposition and does not depend on a self-citation; the point-estimate approximation in Eq. (3) is explicit and, while mathematically uncontrolled as a replacement for the intractable expectation, is not a fitted input disguised as a prediction. The anomaly score is then computed by comparing the input image with the synthesized edited image in CLIP feature space, and quantitative evaluation is performed against held-out ground-truth labels on RUGD and RELLIS. No parameter is fitted to those labels and then reported as a prediction; no load-bearing claim reduces to a self-citation. The only self-citations, such as following Ancha et al. [17] for the normalizing-flow baseline, concern baseline implementation details and do not support the central derivation. The uncontrolled nature of Eq. (3) is a correctness or robustness concern, not a circularity one.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one tuned hyperparameter (guidance strength), a standard-math backbone, and three domain/ad hoc assumptions; no new physical entities are introduced.

free parameters (1)
  • guidance strength λ (or α) = not reported
    Controls the width of the Gaussian similarity kernel rsim(x,y)=exp(-λ||x-y||^2); App. II.C shows the effect of tuning guidance strength on the edited image, and the reported results depend on this choice.
assumptions (4)
  • standard math Bayes rule and standard DDPM score relationships used in Theorem 1
    The proof in App. I re-expresses conditional score functions and applies the first-order Gaussian approximation; this is standard probability and diffusion model math.
  • domain assumption Training images are i.i.d. samples from q(x0), and anomalies are segments whose likelihood under q is low, so a model fitted to q will tend to replace them during editing
    Sec. II defines anomalies implicitly by likelihood under q; the method's success depends on the diffusion model actually erasing OOD objects, which is not guaranteed.
  • ad hoc to paper The point-estimate approximation E_{q(x0|xt)}[rsim] ≈ rsim(µθ_0(xt), xinput) is valid
    Eq. 3 replaces an intractable expectation with the function evaluated at the posterior mean; no error bound is given, and validity is argued only at extreme timesteps.
  • domain assumption CLIP features (MaskCLIP+FeatUp) capture semantically meaningful edits while ignoring pixel-level noise
    Sec. VI relies on cosine distance in CLIP space to separate true edits from global subtle changes; this is an empirical premise validated only through the reported metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Anomalies by Synthesis: Anomaly Detection using Generative Diffusion Models for Off-Road Navigation." pith.science (2026). https://pith.science/paper/QXPEICIB

@misc{pith2026250522805,
  author       = {Pith},
  title        = {Pith review of: Anomalies by Synthesis: Anomaly Detection using Generative Diffusion Models for Off-Road Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXPEICIB}},
  note         = {Machine review of arXiv:2505.22805}
}
read the original abstract

In order to navigate safely and reliably in off-road and unstructured environments, robots must detect anomalies that are out-of-distribution (OOD) with respect to the training data. We present an analysis-by-synthesis approach for pixel-wise anomaly detection without making any assumptions about the nature of OOD data. Given an input image, we use a generative diffusion model to synthesize an edited image that removes anomalies while keeping the remaining image unchanged. Then, we formulate anomaly detection as analyzing which image segments were modified by the diffusion model. We propose a novel inference approach for guided diffusion by analyzing the ideal guidance gradient and deriving a principled approximation that bootstraps the diffusion model to predict guidance gradients. Our editing technique is purely test-time that can be integrated into existing workflows without the need for retraining or fine-tuning. Finally, we use a combination of vision-language foundation models to compare pixels in a learned feature space and detect semantically meaningful edits, enabling accurate anomaly detection for off-road navigation. Project website: https://siddancha.github.io/anomalies-by-diffusion-synthesis/

Figures

Figures reproduced from arXiv: 2505.22805 by the authors.

Figure 1
Figure 1. We present a method for anomaly detection in off-road images using an analysis by synthesis approach. Synthesis: A diffusion model trained on in-distribution data edits the input image to remove out-of-distribution segments. Analysis: Anomaly detection is framed as extracting the difference between input and edited images. The training dataset contains off-road images with: (a-d) natural vegetation, ground and sky, … view at source ↗
Figure 2
Figure 2. Our proposed pipeline for pixel-wise anomaly detection. Left to right: In the synthesis step, a trained diffusion model edits a given input image to remove anomaly segments without modifying other parts of the image. In this case, the model blends the OOD vehicle into dirt in the background. The analysis step extracts anomalies by comparing the pair of images in the CLIP [1] feature space. First, MaskCLIP [2] comput… view at source ↗
Figure 3
Figure 3. Probabilistic graphical model for the conditional forward diffusion process. The target variable we wish to sample is x0. Directed edges correspond to the standard forward diffusion process. The unnormalized factor rsim(x0, x input 0 ) conditions x0 to be similar to the (fixed) input image. III. BACKGROUND ON DIFFUSION MODELS Diffusion models [32, 38, 39] are recently developed, state-of-the-art generative models ef… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Probabilistic graphical model of the conditional forward diffusion process. We wish to sample the random variable x0 corresponding to the training data distribution q(x0). The directed edges between xt−1 and xt (for t = 1, . . . , T) correspond to the vanilla forward d…
Figure 5
Figure 5. Figure 5: Examples of video frames, annotations and semantic classes from the full RUGD dataset [15] [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: In-distribution and out-of-distribution images from the RUGD dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 8
Figure 8. Figure 8: The strength of the diffusion model’s guidance gradient can be tuned by a hyperparameter α. As α decreases, the guidance enforcing similarity between the input image and the image synthesized by the model reduces [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 74 canonical work pages

  1. [1]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. In Proceedings of the International Conference on Machine Learning (ICML) , pages 8748– 8763, 2021

    work page 2021

  2. [2]

    MaskCLIP: Masked self-distillation advances contrastive language-image pretraining

    Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. MaskCLIP: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR) , pages 10995–11005, 2023

    work page 2023

  3. [3]

    Brandt, Axel Feldmann, Zhoutong Zhang, and William T

    Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. FeatUp: A Model-Agnostic Framework for Features at Any Resolution. In Proceedings of the International Conference on Learning Representations (ICLR) , 2024

    work page 2024

  4. [4]

    Segment Anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment Anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015–4026, 2023

    work page 2023

  5. [5]

    Ambler: An autonomous rover for planetary exploration

    John Bares, Martial Hebert, Takeo Kanade, Eric Krotkov, Tom Mitchell, Reid Simmons, and William Whittaker. Ambler: An autonomous rover for planetary exploration. Computer, 22(6):18–26, 1989

    work page 1989

  6. [6]

    Autonomous navigation system for plan- etary exploration rover based on artificial potential fields

    Mauro Massari, Giovanni Giardini, Franco Bernelli- Zazzera, et al. Autonomous navigation system for plan- etary exploration rover based on artificial potential fields. In Proceedings of Dynamics and Control of Systems and Structures in Space (DCSSS) 6th Conference , pages 153–162, 2004

    work page 2004

  7. [7]

    Locomotion policy guided traversability learning using volumetric representations of complex environ- ments

    Jonas Frey, David Hoeller, Shehryar Khattak, and Marco Hutter. Locomotion policy guided traversability learning using volumetric representations of complex environ- ments. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5722–5729. IEEE, 2022

    work page 2022

  8. [8]

    TerrainNet: Visual Modeling of Complex Terrain for High-speed, Off-road Navigation

    Xiangyun Meng, Nathan Hatch, Alexander Lambert, Anqi Li, Nolan Wagener, Matthew Schmittle, JoonHo Lee, Wentao Yuan, Zoey Chen, Samuel Deng, et al. TerrainNet: Visual Modeling of Complex Terrain for High-speed, Off-road Navigation. In Proceedings of Robotics: Science and Systems , Daegu, Republic of Korea, July 2023

    work page 2023

Show all 79 references
  1. [9]

    Semantic terrain classification for off-road autonomous driving

    Amirreza Shaban, Xiangyun Meng, JoonHo Lee, Byron Boots, and Dieter Fox. Semantic terrain classification for off-road autonomous driving. In Conference on Robot Learning, pages 619–629. PMLR, 2022

    2022

  2. [10]

    Visually augmented navigation for autonomous under- water vehicles

    Ryan M Eustice, Oscar Pizarro, and Hanumant Singh. Visually augmented navigation for autonomous under- water vehicles. IEEE Journal of oceanic Engineering , 33(2):103–122, 2008

    2008

  3. [11]

    Vision-Based Goal- Conditioned Policies for Underwater Navigation in the Presence of Obstacles

    Travis Manderson, Juan Camilo Gamboa, Stefan Wap- nick, Jean-Fran c ¸ois Tremblay, Florian Shkurti, Dave Meger, and Gregory Dudek. Vision-Based Goal- Conditioned Policies for Underwater Navigation in the Presence of Obstacles. In Proceedings of Robotics: Science and Systems (R...

    2020

  4. [12]

    Deep multispectral semantic scene understanding of forested environments using multimodal fusion

    Abhinav Valada, Gabriel L Oliveira, Thomas Brox, and Wolfram Burgard. Deep multispectral semantic scene understanding of forested environments using multimodal fusion. In 2016 International Symposium on Experimental Robotics (ISER), pages 465–477. Springer, 2017

    2016

  5. [13]

    AdapNet: Adaptive semantic segmenta- tion in adverse environmental conditions

    Abhinav Valada, Johan Vertens, Ankit Dhall, and Wol- fram Burgard. AdapNet: Adaptive semantic segmenta- tion in adverse environmental conditions. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4644–4651. IEEE, 2017

    2017

  6. [14]

    GA-Nav: Efficient terrain segmen- tation for robot navigation in unstructured outdoor environments

    Tianrui Guan, Divya Kothandaraman, Rohan Chandra, Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, and Dinesh Manocha. GA-Nav: Efficient terrain segmen- tation for robot navigation in unstructured outdoor environments. IEEE Robotics and Automation Letters , 7(3):8138–8145, 2022

    2022

  7. [15]

    A RUGD dataset for autonomous navigation and visual perception in un- structured outdoor environments

    Maggie Wigness, Sungmin Eum, John G Rogers, David Han, and Heesung Kwon. A RUGD dataset for autonomous navigation and visual perception in un- structured outdoor environments. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5000–5007. ...

    2019

  8. [16]

    RELLIS-3D dataset: Data, bench- marks and analysis

    Peng Jiang, Philip Osteen, Maggie Wigness, and Srikanth Saripalli. RELLIS-3D dataset: Data, bench- marks and analysis. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages 1110–1116. IEEE, 2021. URL http://www. unmannedlab.org/research/RELLIS-3D

    2021

  9. [17]

    Osteen, and Nicholas Roy

    Siddharth Ancha, Philip R. Osteen, and Nicholas Roy. Deep Evidential Uncertainty Estimation for Semantic Segmentation under Out-Of-Distribution Obstacles. In IEEE International Conference on Robotics and Automa- tion (ICRA), Yokohama, Japan, May 2024

    2024

  10. [18]

    Prior and Posterior Networks: A Survey on Evidential Deep Learning Methods For Uncertainty Estimation

    Dennis Thomas Ulmer, Christian Hardmeier, and Jes Frellsen. Prior and Posterior Networks: A Survey on Evidential Deep Learning Methods For Uncertainty Estimation. Transactions on Machine Learning Research (TMLR), 2023. ISSN 2835-8856

    2023

  11. [19]

    Posterior network: Uncertainty estimation without OOD samples via density-based pseudo-counts

    Bertrand Charpentier, Daniel Z ¨ugner, and Stephan G¨unnemann. Posterior network: Uncertainty estimation without OOD samples via density-based pseudo-counts. Advances in Neural Information Processing Systems , 33: 1356–1367, 2020

    2020

  12. [20]

    Natural posterior network: Deep Bayesian predictive uncertainty for exponential family distributions

    Bertrand Charpentier, Oliver Borchert, Daniel Z ¨ugner, Simon Geisler, and Stephan G ¨unnemann. Natural posterior network: Deep Bayesian predictive uncertainty for exponential family distributions. In International Conference on Learning Representations , 2022

    2022

  13. [21]

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

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

    2024

  14. [22]

    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

  15. [23]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 , 2016

    2016 arXiv

  16. [24]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems (NeurIPS) , 30, 2017

    2017

  17. [25]

    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 Information Processing Systems , 31, 2018

    2018

  18. [26]

    Concurrent misclassification and out-of-distribution de- tection for semantic segmentation via energy-based nor- malizing flow

    Denis Gudovskiy, Tomoyuki Okuno, and Yohei Nakata. Concurrent misclassification and out-of-distribution de- tection for semantic segmentation via energy-based nor- malizing flow. In Uncertainty in Artificial Intelligence , pages 745–755. PMLR, 2023

    2023

  19. [27]

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

    Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuanhong Chen, and Gustavo Carneiro. Pixel-wise energy-biased abstention learning for anomaly segmenta- tion on complex urban driving scenes. In European Con- ference on Computer Vision , pages 246–263. Springer, 2022

    2022

  20. [28]

    Dense open-set recognition based on training with noisy negative images

    Petra Bevandi ´c, Ivan Kre ˇso, Marin Or ˇsi´c, and Sini ˇsa ˇSegvi´c. Dense open-set recognition based on training with noisy negative images. Image and vision computing, 124:104490, 2022

    2022

  21. [29]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 5128–5137, 2021

    2021

  22. [30]

    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 (ICCV), pages 711–722, 2023

    2023

  23. [31]

    Maskomaly: Zero-shot mask anomaly segmentation

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

    2023 arXiv

  24. [32]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS) , 33:6840– 6851, 2020

    2020

  25. [33]

    Diffusion models in vision: A survey

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023

    2023

  26. [34]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems (NeurIPS) , 34: 8780–8794, 2021

    2021

  27. [35]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages 10684–10695, 2022

    2022

  28. [36]

    Analysis by synthesis: a (re-) emerging program of research for language and vision

    Thomas G Bever and David Poeppel. Analysis by synthesis: a (re-) emerging program of research for language and vision. Biolinguistics, 4(2-3):174–200, 2010

    2010

  29. [37]

    Vision as Bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301–308, 2006

    Alan Yuille and Daniel Kersten. Vision as Bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301–308, 2006

    2006

  30. [38]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning (ICML) , pages 2256–2265. PMLR, 2015

    2015

  31. [39]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score- Based Generative Modeling through Stochastic Differ- ential Equations. In 9th International Conference on Learning Representations (ICLR), 2021, Virtual Event, Austria, M...

    2021

  32. [40]

    Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and MCMC

    Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl- Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and MCMC. In Interna- tional Conferenc...

  33. [41]

    DALL-E: Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. DALL-E: Zero-shot text-to-image generation. In International conference on machine learning (ICML), pages 8821–8831. PMLR, 2021

    2021

  34. [42]

    Dif- fusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Dif- fusion policy: Visuomotor policy learning via action diffusion. In Proceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023

    2023

  35. [43]

    Octo: An open-source generalist robot policy

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Charles Xu, Jianlan Luo, Tobias Kreiman, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo...

    2024

  36. [44]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP) , 13(4):600–612, 2004

    2004

  37. [45]

    DINOv2: Learning Robust Visual Features without Supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. IEEE Transactions on Machine Learning Resea...

    2023

  38. [46]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016

    2016

  39. [47]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at...

    2021

  40. [48]

    Emerging Properties in Self-Supervised Vision Transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) , pages 9650–9660, 2021

    2021

  41. [49]

    Segment- MeIfYouCan: A benchmark for anomaly segmentation

    Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Her- mann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. Segment- MeIfYouCan: A benchmark for anomaly segmentation. In J. Vanschoren and S. Yeung, editors, Proceedings of the International C...

    2021

  42. [50]

    Normalizing flows for probabilistic modeling and inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshmi- narayanan. Normalizing flows for probabilistic modeling and inference. The Journal of Machine Learning Research (JMLR), 22(1):2617–2680, 2021

    2021

  43. [51]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR) , 2021

    2021

  44. [52]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM Computing Surveys (CSUR) , 54(2):1–38, 2021

    2021

  45. [53]

    Robust principal component analysis? Journal of the ACM (JACM) , 58(3):1–37, 2011

    Emmanuel J Cand `es, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? Journal of the ACM (JACM) , 58(3):1–37, 2011

    2011

  46. [54]

    Kernel principal component analysis

    Bernhard Sch ¨olkopf, Alexander Smola, and Klaus- Robert M ¨uller. Kernel principal component analysis. In Proceedings of the International Conference on Artificial Neural Networks (ICANN) , pages 583–588, 1997

    1997

  47. [55]

    Anomaly detection with robust deep autoencoders

    Chong Zhou and Randy C Paffenroth. Anomaly detection with robust deep autoencoders. In Proceedings of the ACM Special-Interest Group on Knowledge Discovery and Data Mining (SIGKDD) International conference, pages 665–674, 2017

    2017

  48. [56]

    Very sparse random projections

    Ping Li, Trevor J Hastie, and Kenneth W Church. Very sparse random projections. In Proceedings of the ACM Special-Interest Group on Knowledge Discovery and Data Mining (SIGKDD) International conference , pages 287–296, 2006

    2006

  49. [57]

    Learning representations of ultrahigh-dimensional data for random distance-based outlier detection

    Guansong Pang, Longbing Cao, Ling Chen, and Huan Liu. Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In Proceedings of the ACM Special-Interest Group on Knowledge Discovery and Data Mining (SIGKDD) International conference, ...

    2018

  50. [58]

    Loda: Lightweight on-line detector of anomalies

    Tom´aˇs Pevn `y. Loda: Lightweight on-line detector of anomalies. Machine Learning, 102:275–304, 2016

    2016

  51. [59]

    A review of novelty detection

    Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal Processing, 99:215–249, 2014

    2014

  52. [60]

    Autonomous navigation in unknown environments using machine learning

    Charles Andrew Richter. Autonomous navigation in unknown environments using machine learning . PhD thesis, Massachusetts Institute of Technology, 2017

    2017

  53. [61]

    High- dimensional and large-scale anomaly detection using a linear one-class svm with deep learning

    Sarah M Erfani, Sutharshan Rajasegarar, Shanika Karunasekera, and Christopher Leckie. High- dimensional and large-scale anomaly detection using a linear one-class svm with deep learning. Pattern Recognition, 58:121–134, 2016

    2016

  54. [62]

    Object-centric auto- encoders and dummy anomalies for abnormal event detection in video

    Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana- Iuliana Georgescu, and Ling Shao. Object-centric auto- encoders and dummy anomalies for abnormal event detection in video. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR), pages 7842–7...

    2019

  55. [63]

    Learning deep representations of appearance and motion for anomalous event detection

    Dan Xu, Jingkuan Song, Yan Yan, E Ricci, Niculae Sebe, et al. Learning deep representations of appearance and motion for anomalous event detection. In Proceedings of the British Machine Vision Conference (BMVC) , pages 1–12, 2015

    2015

  56. [64]

    Netwalk: A flexible deep embedding approach for anomaly detection in dynamic networks

    Wenchao Yu, Wei Cheng, Charu C Aggarwal, Kai Zhang, Haifeng Chen, and Wei Wang. Netwalk: A flexible deep embedding approach for anomaly detection in dynamic networks. In Proceedings of the ACM Special-Interest Group on Knowledge Discovery and Data Mining (SIGKDD) International...

    2018

  57. [65]

    Deep evidential regression

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. Advances in Neural Information Processing Systems , 33:14927– 14937, 2020

    2020

  58. [66]

    EVORA: Deep Evidential Traversability Learning for Risk-Aware Off-Road Autonomy

    Xiaoyi Cai, Siddharth Ancha, Lakshay Sharma, Philip R Osteen, Bernadette Bucher, Stephen Phillips, Jiuguang Wang, Michael Everett, Nicholas Roy, and Jonathan P How. EVORA: Deep Evidential Traversability Learning for Risk-Aware Off-Road Autonomy. IEEE Transactions on Robotics, ...

    2024

  59. [67]

    PIETRA: Physics- Informed Evidential Learning for Traversing Out-of- Distribution Terrain

    Xiaoyi Cai, James Queeney, Tong Xu, Aniket Datar, Chenhui Pan, Max Miller, Ashton Flather, Philip R Os- teen, Nicholas Roy, Xuesu Xiao, et al. PIETRA: Physics- Informed Evidential Learning for Traversing Out-of- Distribution Terrain. arXiv preprint arXiv:2409.03005 , 2024

    2024 arXiv

  60. [68]

    Uncertainty-aware panoptic segmentation

    Kshitij Sirohi, Sajad Marvi, Daniel B ¨uscher, and Wol- fram Burgard. Uncertainty-aware panoptic segmentation. IEEE Robotics and Automation Letters (RAL), 8(5):2629– 2636, 2023

    2023

  61. [69]

    Re- construction by inpainting for visual anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. Re- construction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021

    2021

  62. [70]

    Learning deep representations of appearance and motion for anomalous event detection

    Xu Dan, Elisa Ricci, Yan Yan, Jingkuan Song, Nicu Sebe, et al. Learning deep representations of appearance and motion for anomalous event detection. In Proceed- ings of the British Machine Vision Conference (BMVC) , pages 8–1, 2015

    2015

  63. [71]

    Unsupervised deep anomaly detection in chest radiographs

    Takahiro Nakao, Shouhei Hanaoka, Yukihiro No- mura, Masaki Murata, Tomomi Takenaga, Soichiro Miki, Takeyuki Watadani, Takeharu Yoshikawa, Naoto Hayashi, and Osamu Abe. Unsupervised deep anomaly detection in chest radiographs. Journal of Digital Imaging, 34:418–427, 2021

    2021

  64. [72]

    Unsu- pervised anomaly detection with generative adversarial networks to guide marker discovery

    Thomas Schlegl, Philipp Seeb ¨ock, Sebastian M Wald- stein, Ursula Schmidt-Erfurth, and Georg Langs. Unsu- pervised anomaly detection with generative adversarial networks to guide marker discovery. In Proceedings of the International Conference on Information Processing in Med...

    2017

  65. [73]

    f- anogan: Fast unsupervised anomaly detection with gen- erative adversarial networks

    Thomas Schlegl, Philipp Seeb ¨ock, Sebastian M Wald- stein, Georg Langs, and Ursula Schmidt-Erfurth. f- anogan: Fast unsupervised anomaly detection with gen- erative adversarial networks. Medical Image Analysis , 54:30–44, 2019

    2019

  66. [74]

    Ganomaly: Semi-supervised anomaly detection via adversarial training

    Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 622–637, 2019

    2019

  67. [75]

    Skip-ganomaly: Skip connected and adver- sarially trained encoder-decoder anomaly detection

    Samet Akc ¸ay, Amir Atapour-Abarghouei, and Toby P Breckon. Skip-ganomaly: Skip connected and adver- sarially trained encoder-decoder anomaly detection. In Proceedings of the IEEE International Joint Conference on Neural Networks (IJCNN) , pages 1–8, 2019

    2019

  68. [76]

    Diffusion models for medical anomaly detection

    Julia Wolleb, Florentin Bieder, Robin Sandk ¨uhler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 35–45. Springer, 2022

    2022

  69. [77]

    Fast unsupervised brain anomaly detection and segmentation with diffusion models

    Walter HL Pinaya, Mark S Graham, Robert Gray, Pedro F Da Costa, Petru-Daniel Tudosiu, Paul Wright, Yee H Mah, Andrew D MacKinnon, James T Teo, Rolf Jager, et al. Fast unsupervised brain anomaly detection and segmentation with diffusion models. In International Conference on Me...

    2022

  70. [78]

    DiffusionAD: Norm-guided one-step denois- ing diffusion for anomaly detection

    Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. DiffusionAD: Norm-guided one-step denois- ing diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023

    2023 arXiv

  71. [79]

    obstacle

    Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly Detection with Conditioned Denoising Diffu- sion Models. arXiv preprint arXiv:2305.15956 , 2023. Appendix Anomalies-by-Synthesis: Anomaly Detection using Generative Diffusion Models for Off-Road Navigation Sunshine Jiang ...

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.