Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection

Aihua Zheng; Jian Liang; Ran He; Shuo Lu; Yingsheng Wang

REVIEW 4 major objections 4 minor 51 references

Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ClaFR claims OOD detection from the classifier's weight matrix alone: SVD top singular vectors define a subspace, and the projection norm separates ID from OOD samples.

desk verdict A simple, training-data-free OOD detector built on SVD of the classifier weights that delivers strong benchmark numbers, but the load-bearing geometric premise is asserted rather than proven. read the letter →

arxiv 2509.06988 v1 pith:ETF6V44S submitted 2025-09-02 cs.CV cs.AIcs.LG

Yingsheng Wang , Shuo Lu , Jian Liang , Aihua Zheng , Ran He This is my paper

classification cs.CVcs.AIcs.LG

keywords Out-of-DistributionDetectionSubspaceProjectionFeatureReconstructionSingularValueDecompositionClassifierWeightsPost-hocTraining-freeImageNet

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 introduces ClaFR, a post-hoc out-of-distribution detection method that uses only the linear classifier's weight matrix and the penultimate-layer features at inference. It performs SVD on the weight matrix, keeps the top singular vectors that account for about 90% of the singular-value energy, and scores each sample by the norm of its feature projection onto that subspace. The claim is that this class-known subspace captures the same dominant directions as the in-distribution features, so ID samples project with larger norm while OOD samples lose more information. If correct, OOD detection no longer needs the training set, which matters for data-privacy and fast model-update scenarios; the method also runs in O(1) per sample, far cheaper than nearest-neighbor or stored-feature baselines. On ImageNet-1k and CIFAR benchmarks, the paper reports the best average AUROC/FPR95 among the compared post-hoc methods.

What carries the argument

The class-known subspace U_M: the top-m left singular vectors of the classifier weight matrix W, obtained by SVD, where m is chosen so the retained singular values sum to roughly 90% of the total. The paper treats U_M as a proxy for the principal-component subspace of in-distribution features. The score is the norm of the feature projection onto U_M; because the SVD columns are orthonormal, the feature reconstruction error -||z U_M U_M^T - z||_2 simplifies to S(x)=||z U_M||_2, a single lightweight projection computed at inference.

What would settle it

Take a trained model and compute both the top singular vectors of its classifier weight matrix and the principal components of its in-distribution training features. If the subspace angle between these sets is large, or if the AUROC of S(x)=||z U_M||_2 does not drop when W is replaced by a scrambled-weight matrix while features are kept fixed, the central claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a subspace read directly off the classifier weights—not from training features—is enough to separate in-distribution from out-of-distribution data. By SVD-decomposing W = U Σ V^T and taking the top m left singular vectors U_M (m chosen so the retained singular values sum to about 90% of the total), ClaFR defines a 'class-known subspace.' The OOD score is S(x)=||z U_M||_2; equivalently, it is the negative feature reconstruction error after projecting z onto U_M and back. The paper reports that this score outperforms ten post-hoc baselines in average AUROC on ImageNet-1k with ResNet-50 (89.32%) and MobileNet (83.63%), and is competitive o

Load-bearing premise

The load-bearing premise is that the top singular vectors of the classifier weight matrix span the same subspace as the principal components of the in-distribution feature distribution, so in-distribution features have larger projection norms onto U_M than out-of-distribution features; the paper asserts this without derivation or verification.

Editorial extensions

If this is right

OOD detection can be performed using only the classifier weights, so deployment in privacy-sensitive settings no longer requires shipping or storing training features.
When the model is updated or new classes are added, the detector is refreshed by re-running SVD on the new weight matrix—no retraining over the database.
The O(1) per-sample cost and roughly 12 MiB storage make the method practical for edge or large-scale use, where KNN-style baselines need O(N_tr) search and around 20 GiB.
The score is a pure geometric quantity built from the model's own parameters, so it transfers across backbones such as ResNet-50, MobileNet, and ResNet-34 without dataset-specific tuning.

Reading between the lines

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

The paper's premise—that top singular vectors of W align with the principal components of ID features—is untested; directly measuring this alignment on standard models would show how general the result is beyond the contrastively trained backbones used here.
If the weight matrix alone encodes feature geometry, the same subspace idea might extend to vision transformers or models without a conventional linear classifier by using a substitute weight-like operator.
Because the score is a projection norm, OOD samples whose features happen to align with dominant ID directions could evade detection; a singular-value-weighted or normalized variant is a natural extension not explored in 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.

Desk Editor's Note

read the letter

You should know three things about this paper. First, it proposes ClaFR, an OOD scoring function that uses only the classifier weight matrix: SVD of W gives a subspace, and the score is the projection norm of the normalized penultimate feature onto that subspace. No training data is needed at inference. Second, the empirical results are genuinely good on the reported benchmarks — top average AUROC/FPR95 on ImageNet-1k and competitive on CIFAR, with a huge efficiency win over KNN. Third, the paper's central theoretical justification is unverified, and the presentation has a few internal inconsistencies that need cleaning up. What's new: using the weight matrix's singular vectors as a substitute for PCA on training features is a simple and practical trick. It is close in spirit to ViM and PCA-reconstruction, but explicitly avoids touching the training set, which matters for privacy-sensitive deployments. The authors also release code, which makes the method reproducible. That is real value. Where it gets soft. The paper claims the top singular vectors of W span the same subspace as the principal components of ID features, but this is asserted, not derived. If that alignment fails, the score has no reason to separate ID from OOD. The empirical success suggests the alignment is at least partially true, likely because supervised contrastive training makes class prototypes and ID feature directions coincide, but the paper doesn't test or demonstrate this. A simple experiment comparing ClaFR's subspace to PCA on training features would have addressed the concern; its absence is the main weakness. There are also technical sloppinesses. Eq. (3) calls e(x) a reconstruction error, then Eq. (4) claims it's 'another representation' of the same thing. After the Algorithm 1 normalization, the two are monotonically related, so the ranking is valid, but that's not what the text says. The normalization itself is ambiguous: Algorithm 1 normalizes z to unit norm before computing the score, while Eq. (3) uses raw z. The CIFAR table reports only averages, no per-dataset numbers or error bars, which makes it hard to judge variance. The hyperparameter alpha=0.9 is selected on the OOD benchmarks; their ablation shows a broad plateau, so this is a minor concern, but worth stating. Overall, the central empirical claim holds up on the reported benchmarks, but the theoretical premise needs support. This is a solid conference paper in need of revision, not a desk reject. The idea is simple enough that it will be cited and built upon regardless, so referees should focus on whether the subspace-alignment assumption can be verified or at least rigorously motivated. I would send it to peer review with a request for stronger analysis and fuller results reporting.

Referee Report

4 major / 4 minor

Summary. The paper proposes ClaFR, a post-hoc, training-data-free OOD detection method that computes an OOD score by projecting penultimate-layer features onto a subspace spanned by the top left singular vectors of the classifier weight matrix W. The subspace dimension is selected by a cumulative singular-value ratio alpha ~ 0.9. The score is S(x) = ||z U_M||_2, intended as a proxy for negative feature reconstruction error. Experiments on ImageNet-1k (ResNet-50, MobileNet) and CIFAR-10/100 (ResNet-34) report competitive or state-of-the-art average AUROC/FPR95 against ten baselines, along with a complexity comparison showing O(1) inference and negligible storage. The central claim is that ClaFR does not need access to training data while achieving leading performance.

Significance. If the reported results hold, ClaFR is a practically attractive OOD detector: it uses only the classifier weights, has near-zero computational overhead, requires no training-data access, and is simple to implement. The paper's strengths include released code, a clean algorithmic idea, and competitive numbers on large-scale benchmarks. However, the significance is tempered by the lack of a validated geometric justification. The method's discriminative power depends on the unverified premise that W's top singular vectors align with the principal directions of ID features; without this, the method is just a heuristic that happens to work on the tested benchmarks. The paper would be substantially strengthened by a diagnostic linking U_M to the ID feature covariance and by clarifying the score's mathematical relationship to reconstruction error.

major comments (4)

[Sec. 4.2, Eq. (3) vs. Eq. (4)] The paper states that Eq. (4) is "another representation of the feature reconstruction error", but the two expressions are not equivalent as written. Eq. (3) defines e(x) = -||z U_M U_M^T - z||_2 = -sqrt(||z||^2 - ||z U_M||^2). This is a monotone transform of S(x) = ||z U_M|| only when ||z|| is constant. Algorithm 1 normalizes z to unit norm, but Eq. (3) uses raw z; the text does not clarify which z is used in the experiments. If normalization is applied, the score is a direction-based (cosine-like) measure, and the reconstruction-error interpretation changes. Please define the feature normalization unambiguously and state the exact monotone relationship that justifies using S(x) as the score.
[Sec. 4.1, statement that "UM is regarded as the principal component of ID data features"] This is the load-bearing geometric premise, and it is only asserted, not derived or verified. The classifier weight matrix W is learned to separate classes; under supervised contrastive training, W encodes class prototypes or decision boundaries, not necessarily the top principal components of the ID feature distribution. If W's singular vectors do not align with the high-variance ID feature directions, the claimed separation between ID and OOD is unsupported. Please provide a theoretical justification or, at minimum, an empirical diagnostic (e.g., cosine similarity between U_M and the PCA directions of training features, or an ablation with a random orthonormal subspace of the same dimension). Without this, the method's mechanism is not established.
[Algorithm 1, line defining m] The pseudo-code line "m <- P_m i=1 sigma_i > alpha P_D i=1 sigma_i" is not syntactically meaningful. It should read: choose the smallest m such that the cumulative sum of the top-m singular values exceeds alpha times the total sum. Also, the algorithm normalizes z_hat <- z / ||z||_2 but then returns S(x) = ||z U_M||_2 using the unnormalized z. This ambiguity directly affects the score and the reconstruction-error equivalence in Eq. (3). Please correct the pseudo-code and state which feature (raw or normalized) is used in each equation and in the reported experiments.
[Sec. 5.1, Table 2 and CIFAR results] The paper reports only average CIFAR results, with the per-dataset numbers deferred to "the supplementary materials", which are not included in the arXiv submission. Averages can hide large per-dataset variability; for a method that claims "leading performance", the per-OOD-dataset breakdown (CIFAR-10 vs. CIFAR-100, SVHN) is essential for assessing robustness. Additionally, no variance or error bars are reported anywhere, so it is unclear whether the average improvements over the second-best method are within noise. Please include per-dataset tables and, ideally, multiple-seed standard deviations.

minor comments (4)

[Abstract] The sentence "This may not be suitable in scenarios where data privacy protection is a concern" is accidentally duplicated. Please remove the duplicate.
[Sec. 5.1, baselines] The text says "including five output-based methods: ... and six feature-based methods", but only four output-based methods are listed (MSP, Energy, Maxlogit, KL-Matching). Please correct the count or the list.
[References] There are several typographical errors in the reference list, e.g., "Proc. NeruIPS" (should be NeurIPS), "Pro. CVPR" (should be Proc. CVPR), and inconsistent formatting of author names. A careful proofread is needed.
[Sec. 5.4, Table 3] The complexity comparison with KNN is informative, but the "O(1)" claim should be qualified: the OOD score computation is O(D*m) per sample, where D and m are feature dimension and subspace dimension, not strictly O(1) in the feature dimension. Please specify the exact per-sample complexity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ClaFR's OOD score is a deterministic function of classifier weights and test features; the central geometric premise is assumed rather than derived, but it is not input-equivalent.

full rationale

The paper does not exhibit a derivation that reduces to its own inputs. The OOD score S(x)=||zU_M||_2 in Eq. (4) is computed deterministically from the classifier weight matrix W (via SVD) and the test feature z; no parameter is fitted to OOD labels and no OOD benchmark data is used to construct U_M. The claimed equivalence between the reconstruction error in Eq. (3) and the projection norm in Eq. (4) is a monotone transform once Algorithm 1's unit-norm normalization is applied (||z||=1), so it is not a definitional smuggling of the conclusion. The central premise that U_M spans the same subspace as ID feature principal components is asserted rather than proved — Section 4.1 states 'UM is regarded as the principal component of ID data features' — but an unsupported geometric assumption is a correctness risk, not circularity: the method's outputs are not definitionally equal to that premise. The cited prototype result [41] and the PCA-reconstruction works [11,12] are external, and they are not invoked as a self-referential uniqueness theorem. Self-citations [21,26,47] appear only in background context and are not load-bearing. The only hyperparameter α is chosen by cumulative explained variance and is ablated over a range (80%–95%) with stable performance; the paper does not report benchmark-specific fitting of α to OOD labels. Therefore no specific circular reduction can be identified.

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

The method introduces no new entities. It depends on the geometric assumption that the classifier weight matrix's top singular subspace captures ID structure, on the hand-chosen threshold alpha, and on standard SVD. No fitting to OOD labels occurs, so circularity burden is low.

free parameters (1)

Cumulative explained variance ratio alpha = 0.9
Hand-set threshold for choosing the subspace dimension m; validated on OOD benchmarks in ablation (robust between 0.8 and 0.95). Not fitted to OOD labels, but tuned on the test OOD sets.

assumptions (4)

standard math The weight matrix W admits an SVD W = U Sigma V^T with U orthogonal.
Used in Eq. (2); standard linear algebra.
domain assumption The top singular vectors of W span the same subspace as the principal components of ID features.
Core geometric premise in Section 4.1; if false, the projection norm cannot separate ID from OOD.
domain assumption ID feature projections onto the class-known subspace have larger norm than OOD feature projections.
Discriminative premise; the entire score is based on this empirical separation, asserted without formal proof.
domain assumption The 90% singular-value threshold preserves ID-relevant information across different backbones and datasets.
The same alpha is used for all experiments; ablation shows broad robustness, but it is an empirical choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/ETF6V44S

@misc{pith2026250906988,
  author       = {Pith},
  title        = {Pith review of: Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETF6V44S}},
  note         = {Machine review of arXiv:2509.06988}
}

read the original abstract

Out-of-distribution (OOD) detection helps models identify data outside the training categories, crucial for security applications. While feature-based post-hoc methods address this by evaluating data differences in the feature space without changing network parameters, they often require access to training data, which may not be suitable for some data privacy scenarios. This may not be suitable in scenarios where data privacy protection is a concern. In this paper, we propose a simple yet effective post-hoc method, termed Classifier-based Feature Reconstruction (ClaFR), from the perspective of subspace projection. It first performs an orthogonal decomposition of the classifier's weights to extract the class-known subspace, then maps the original data features into this subspace to obtain new data representations. Subsequently, the OOD score is determined by calculating the feature reconstruction error of the data within the subspace. Compared to existing OOD detection algorithms, our method does not require access to training data while achieving leading performance on multiple OOD benchmarks. Our code is released at https://github.com/Aie0923/ClaFR.

Figures

Figures reproduced from arXiv: 2509.06988 by the authors.

**Figure 1.** The overview of ClaFR. During inference, after extracting features through the pre-trained network, these features are projected into the subspace obtained by performing SVD decomposition on the classifier weights. Each sample’s OOD score is calculated by (4). Generally, the lower the score, the more likely the sample is considered to be OOD. Although training-agnostic methods do not require modifications to the or… view at source ↗

**Figure 2.** An ablation study on the sensitivity of the cumulative explained variance ratio α is conducted to veri its impact on OOD detection performance. Our experimental results indicate that the hyperparameter α yields competitive results over a wide range of values, suggesting that ClaFR is not overly sensitive to the choice of α. When the selected proportion of singular values α is approximately 80% to 95%, the model cons… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 49 canonical work pages

[1]

In: Proc

Ammar, M.B., Belkhir, N., Popescu, S., Manzanera, A., Franchi, G.: Neco: Neural collapse based out-of-distribution detection. In: Proc. ICLR (2024)

work page 2024
[2]

In: Proc

Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: Proc. ICML (2022)

work page 2022
[3]

In: Proceedings of the ieee/cvf inter- national conference on computer vision

Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf inter- national conference on computer vision. pp. 5128–5137 (2021)

work page 2021
[4]

In: Proc

Chen, Y.H., Chen, W.Y., Chen, Y.T., Tsai, B.C., Frank Wang, Y.C., Sun, M.: No more discrimination: Cross city adaptation of road scene segmenters. In: Proc. ICCV. pp. 1992– 2001 (2017)

work page 1992
[5]

In: Proc

Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR. pp. 3606–3613 (2014)

work page 2014
[6]

Outlier Detection through Null Space Analysis of Neural Networks

Cook, M., Zare, A., Gader, P.: Outlier detection through null space analysis of neural networks. arXiv preprint arXiv:2007.01263 (2020)

work page Pith review arXiv 2007
[7]

In: Nature Commu- nications (2020)

Daniel, C.d.C., Ian, W., Ben, G.: Causality matters in medical imaging. In: Nature Commu- nications (2020)

work page 2020
[8]

In: Proc

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR. pp. 248–255. Ieee (2009)

work page 2009

Show all 51 references

[9]

In: Proc

Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out- of-distribution detection. In: Proc. ICLR (2023)

2023
[10]

In: Proc

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., De- hghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. ICLR (2020)

2020
[11]

In: Proc

Fang, K., Tao, Q., Lv, K., He, M., Huang, X., Yang, J.: Kernel pca for out-of-distribution detection. In: Proc. NeruIPS (2024)

2024
[12]

In: Proc

Guan, X., Liu, Z., Zheng, W.S., Zhou, Y., Wang, R.: Revisit pca-based technique for out-of- distribution detection. In: Proc. ICCV. pp. 19431–19439 (2023)

2023
[13]

In: Proc

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. CVPR. pp. 770–778 (2016)

2016
[14]

In: Proc

He, R., Hu, B., Zheng, W.S., Guo, Y.: Two-stage sparse representation for robust recognition on large-scale database. In: Proc. AAAI. vol. 24, pp. 475–480 (2010)

2010
[15]

IEEE Transactions on Image Processing 24(12), 5543–5556 (2015)

He, R., Zhang, M., Wang, L., Ji, Y., Yin, Q.: Cross-modal subspace learning via pairwise constraints. IEEE Transactions on Image Processing 24(12), 5543–5556 (2015)

2015
[16]

Hein, M., Andriushchenko, M., Bitterwolf, J.: Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In: Pro. CVPR. pp. 41–50 (2019)

2019
[17]

In: Proc

Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: Proc. ICLR (2017)

2017
[18]

In: Proc

Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. In: Proc. NeurIPS. vol. 34, pp. 677–689 (2021) Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection 13

2021
[19]

Handbook of Systemic Autoimmune Diseases 1(4) (2009)

Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Handbook of Systemic Autoimmune Diseases 1(4) (2009)

2009
[20]

In: Proc

Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In: Proc. NeurIPS. vol. 31 (2018)

2018
[21]

In: Proc

Liang, J., Sheng, L., Wang, Z., He, R., Tan, T.: Realistic unsupervised clip fine-tuning with universal entropy optimization. In: Proc. ICML (2024)

2024
[22]

In: Proc

Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proc. ICLR (2018)

2018
[23]

In: Proc

Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. In: Proc. NeurIPS. vol. 33, pp. 21464–21475 (2020)

2020
[24]

IEEE Transactions on Medical Imaging41(7), 1897–1908 (2022)

Liu, X., Yuan, Y.: A source-free domain adaptive polyp detection framework with style diversification flow. IEEE Transactions on Medical Imaging41(7), 1897–1908 (2022)

1908
[25]

In: Proc

Lu, H., Gong, D., Wang, S., Xue, J., Yao, L., Moore, K.: Learning with mixture of prototypes for out-of-distribution detection. In: Proc. ICLR (2024)

2024
[26]

ACM Comput

Lu, S., Wang, Y., Sheng, L., He, L., Zheng, A., Liang, J.: Out-of-distribution de- tection: A task-oriented survey of recent advances. ACM Comput. Surv. (Aug 2025). https://doi.org/10.1145/3760390

2025 doi
[27]

ICLR (2023)

Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of- distribution detection? In: Proc. ICLR (2023)

2023
[28]

In: Proc

Morteza, P., Li, Y.: Provable guarantees for understanding out-of-distribution detection. In: Proc. AAAI. vol. 36, pp. 7831–7840 (2022)

2022
[29]

In: Proc

Mueller, M., Hein, M.: Mahalanobis++: Improving ood detection via feature normalization. In: Proc. ICML (2025)

2025
[30]

In: Proc

Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y., et al.: Reading digits in natural images with unsupervised feature learning. In: Proc. NeurIPS. vol. 2011, p. 7. Granada, Spain (2011)

2011
[31]

In: Proc

Park, J., Jung, Y.G., Teoh, A.B.J.: Nearest neighbor guidance for out-of-distribution detection. In: Proc. ICCV. pp. 1686–1695 (2023)

2023
[32]

Peng, B., Luo, Y., Zhang, Y., Li, Y., Fang, Z.: Conjnorm: Tractable density estimation for out-of-distribution detection. Proc. ICLR (2024)

2024
[33]

In: Proc

Pu, Y., Feng, Z., Wang, Z., Yang, Z., Li, J.: Anomaly detection for in situ marine plankton images. In: Proc. ICCV. pp. 3661–3671 (2021)

2021
[34]

In: Proc

Qu, S., Zou, T., He, L., R ¨ohrbein, F., Knoll, A., Chen, G., Jiang, C.: Lead: Learning de- composition for source-free universal domain adaptation. In: Proc. CVPR. pp. 23334–23343 (2024)

2024
[35]

In: Proc

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted resid- uals and linear bottlenecks. In: Proc. CVPR. pp. 4510–4520 (2018)

2018
[36]

In: Proc

Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. In: Proc. ICLR (2021)

2021
[37]

In: Multimedia systems

Shan, L., Yichao, T., Ying, T., Hansong, S.: Visual driving assistance system based on few-shot learning. In: Multimedia systems. pp. 2853–2863 (2021)

2021
[38]

In: Proc

Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. In: Proc. NeurIPS (2021)

2021
[39]

In: Proc

Sun, Y., Li, Y.: Dice: Leveraging sparsification for out-of-distribution detection. In: Proc. ECCV. pp. 691–708. Springer (2022)

2022
[40]

In: Proc

Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: Proc. ICML. pp. 20827–20840. PMLR (2022)

2022
[41]

Tanwisuth, K., Fan, X., Zheng, H., Zhang, S., Zhang, H., Chen, B., Zhou, M.: A prototype- oriented framework for unsupervised domain adaptation. Proc. NeurIPS 34, 17194–17208 (2021) 14 Y. Wang et al

2021
[42]

In: Proc

Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proc. CVPR. pp. 8769–8778 (2018)

2018
[43]

In: Proc

Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proc. CVPR. pp. 4921–4930 (2022)

2022
[44]

In: Proc

Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: Proc. CVPR. pp. 3485–3492. IEEE (2010)

2010
[45]

In: Proc

Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. In: Proc. NeurIPS. vol. 35, pp. 32598–32611 (2022)

2022
[46]

In: IJCV

Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. In: IJCV. vol. 132, pp. 5635–5662. Springer (2024)

2024
[47]

arXiv preprint arXiv:2303.12267 (2023)

Yang, P., Liang, J., Cao, J., He, R.: Auto: Adaptive outlier optimization for online test-time ood detection. arXiv preprint arXiv:2303.12267 (2023)

2023 arXiv
[48]

arXiv preprint arXiv:2306.09301 (2023)

Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Li, Y., Liu, Z., et al.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023)

2023 arXiv
[49]

Zhao, Q., Xu, M., Gupta, K., Asthana, A., Zheng, L., Gould, S.: Towards optimal feature- shaping methods for out-of-distribution detection. Proc. ICLR (2024)

2024
[50]

IEEE TPAMI40(6), 1452–1464 (2017)

Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE TPAMI40(6), 1452–1464 (2017)

2017
[51]

In: NeurIPS

Zhu, Y., Chen, Y., Xie, C., Li, X., Zhang, R., Xue, H., Tian, X., Chen, Y., et al.: Boosting out-of-distribution detection with typical features. In: NeurIPS. vol. 35, pp. 20758–20769 (2022)

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

In: Proc

Ammar, M.B., Belkhir, N., Popescu, S., Manzanera, A., Franchi, G.: Neco: Neural collapse based out-of-distribution detection. In: Proc. ICLR (2024)

work page 2024

[2] [2]

In: Proc

Basart, S., Mantas, M., Mohammadreza, M., Jacob, S., Dawn, S.: Scaling out-of-distribution detection for real-world settings. In: Proc. ICML (2022)

work page 2022

[3] [3]

In: Proceedings of the ieee/cvf inter- national conference on computer vision

Chan, R., Rottmann, M., Gottschalk, H.: Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In: Proceedings of the ieee/cvf inter- national conference on computer vision. pp. 5128–5137 (2021)

work page 2021

[4] [4]

In: Proc

Chen, Y.H., Chen, W.Y., Chen, Y.T., Tsai, B.C., Frank Wang, Y.C., Sun, M.: No more discrimination: Cross city adaptation of road scene segmenters. In: Proc. ICCV. pp. 1992– 2001 (2017)

work page 1992

[5] [5]

In: Proc

Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proc. CVPR. pp. 3606–3613 (2014)

work page 2014

[6] [6]

Outlier Detection through Null Space Analysis of Neural Networks

Cook, M., Zare, A., Gader, P.: Outlier detection through null space analysis of neural networks. arXiv preprint arXiv:2007.01263 (2020)

work page Pith review arXiv 2007

[7] [7]

In: Nature Commu- nications (2020)

Daniel, C.d.C., Ian, W., Ben, G.: Causality matters in medical imaging. In: Nature Commu- nications (2020)

work page 2020

[8] [8]

In: Proc

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. CVPR. pp. 248–255. Ieee (2009)

work page 2009