pith. sign in

arxiv: 2606.24404 · v1 · pith:B46KUX3Inew · submitted 2026-06-23 · 💻 cs.CV

Modality-Aware Out-of-Distribution Detection for Multi-Modal Action Recognition

Lars Doorenbos , Duc Manh Vu , Serdar Ozsoy , Juergen Gall This is my paper

Pith reviewed 2026-06-26 00:27 UTC · model grok-4.3

classification 💻 cs.CV
keywords multi-modal OOD detectionaction recognitionpost-hoc detectoruni-modal predictionsfeature-space scoreMultiOOD benchmarkhybrid detectormodality-aware detection
0
0 comments X

The pith

Multi-modal action recognition gains a stronger OOD detector by contrasting full-model predictions against single-modality branches.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that the difference between a multi-modal model's output and the outputs of its individual modality components supplies a reliable signal for identifying out-of-distribution samples. The authors construct a post-hoc detector around this signal, merge it with a feature-space score that flags off-manifold points, and normalize the result by the multi-modal logits. The hybrid detector requires no retraining, works alongside existing training-time regularization methods, and raises average detection performance across the MultiOOD benchmark datasets. A reader would care because current multi-modal systems largely reuse uni-modal OOD detectors at inference and therefore leave modality-specific robustness information unused.

Core claim

Based on an observed relationship between multi-modal and uni-modal predictions, we propose a post-hoc detector that combines this signal with a feature-space score and normalizes the combination by multi-modal logits; the resulting hybrid detector is compatible with training-time approaches and outperforms the state of the art on average across established datasets from the MultiOOD benchmark, showing the value of explicitly considering different modalities at inference time.

What carries the argument

The relationship between multi-modal and uni-modal predictions, used as an explicit signal and combined with a normalized feature-space score to form a hybrid post-hoc OOD detector.

If this is right

  • The detector can be paired directly with any existing training-time OOD regularization method without modification.
  • Average OOD detection performance rises across a range of established multi-modal action recognition datasets.
  • Explicit use of modality-specific predictions at inference time improves robustness beyond what uni-modal detectors achieve.
  • Normalization by multi-modal logits preserves the prediction-gap signal while avoiding new biases in the score.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prediction-gap idea could be tested in other multi-modal settings such as audio-visual or vision-language models to check whether modality contrast remains useful.
  • Deployed systems that combine video, audio, and sensor streams might reduce missed OOD events by adding this lightweight contrast at test time.
  • A direct follow-up experiment would replace the logit normalization with alternative scaling factors and measure whether detection margins change.

Load-bearing premise

The gap between multi-modal and uni-modal predictions stays informative and stable for separating in-distribution from out-of-distribution samples across models and datasets.

What would settle it

If the hybrid detector shows no consistent gain over standard uni-modal OOD detectors when evaluated on additional multi-modal action datasets where the uni-modal branches are forced to produce identical outputs to the full model, the claimed utility of the prediction-gap signal would be refuted.

Figures

Figures reproduced from arXiv: 2606.24404 by Duc Manh Vu, Juergen Gall, Lars Doorenbos, Serdar Ozsoy.

Figure 1
Figure 1. Figure 1: Motivation. a) Current methods for multi-modal OOD detection apply off￾the-shelf detectors designed for the uni-modal case at inference time. b) In contrast, we propose a modality-aware OOD detector that improves multi-modal OOD detection. OOD detection [16, 42] methods learn scoring functions that measure the level of out-of-distributionness of test samples, such that they can be filtered out. However, mo… view at source ↗
Figure 2
Figure 2. Figure 2: Example of our observed relation between uni- and multi-modal pre [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Finding the invariants on a toy example. The principal com￾ponents with the lowest variance de￾scribe off-manifold directions in the multi-modal feature space. We use these to detect feature-level OOD samples. Any deviation in these tight dimensions is a strong signal that a sample lies off￾manifold and is OOD. Furthermore, devi￾ations in dimensions with smaller variance should have a larger impact than si… view at source ↗
Figure 4
Figure 4. Figure 4: Despite being trained with A2D, the difference between uni￾modal predictions does not provide a strong indicator for OOD. Instead, we find that the relation between the pre￾dictions of the uni-modal heads and the prediction of the multi-modal head is a strong signal of normality. two modalities can alleviate some of these shortcomings. The combination of s(x) and g(x), for example, reaches the second-best … view at source ↗
Figure 5
Figure 5. Figure 5: Ablation studies for our framework. We evaluate the impact of (a) the distance function, (b) the choice for probabilities over logits in Eq. (4), and (c) variants of feature-based scores for r(x). 0.0 0.25 0.4 0.5 0.75 1.0 1.25 s 90 92 94 96 98 100 102 A U C Near-OOD 0.0 0.25 0.4 0.5 0.75 1.0 1.25 r 90 92 94 96 98 100 102 A U C Near-OOD 0.2 0.5 1.0 2.0 5.0 p 90 92 94 96 98 100 102 A U C Near-OOD (a) (b) (c… view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity to hyperparameters. We show the AUC for HMDB51:Kinetics Far-OOD for different settings of the hyperparameters a) γs, b) γr, and c) p. Our method is robust to their settings. fully connected W ∈ R (MC)×C (“fully connected"), where every uni-modal logit influences every multi-modal logit, and a version with only a single w˜ shared between all classes (“single linear"). We show the results in Tab.… view at source ↗
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity on Kinetics Near-OOD. Our method is robust to a) γs, b) γr, and c) p, showing that it does not require task-specific tuning. Ours: 95.5 DPUf: 86.2 DPUp: 97.5 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: HMDB Near-OOD sample confidently mislabeled as ID by all three [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: HMDB Near-OOD sample only detected confidently by our method. [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: HMDB ID sample detected as mostly OOD by all methods. [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: HMDB:UCF Far-OOD sample confidently detected as OOD by all [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
read the original abstract

The incorporation of additional modalities into action recognition models increases their performance across a wide range of settings. However, how this additional information can contribute to making the models more robust remains underexplored, particularly for the case of multi-modal out-of-distribution (OOD) detection. While methods exist that regularize the multi-modal training process with OOD detection in mind, they still apply off-the-shelf OOD detectors designed for the uni-modal case during inference, discarding important information. Based on an interesting relationship we find between the multi-modal and uni-modal predictions, we propose to use this signal to build a post-hoc detector explicitly designed for the multi-modal scenario. We combine this new source of information with a feature-space score, which detects off-manifold samples in the multi-modal space, and normalize them by the multi-modal logits. In doing so, the proposed hybrid detector is compatible with existing training-time approaches and consistently improves performance. Experiments on a wide range of established datasets from the MultiOOD benchmark show that, on average, our approach outperforms the state of the art. Our results show the importance of explicitly considering the different modalities at inference time for multi-modal OOD detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a post-hoc, modality-aware OOD detector for multi-modal action recognition. It identifies an empirical relationship between multi-modal and uni-modal predictions, combines this signal with a feature-space score for off-manifold detection, and normalizes the result by multi-modal logits. The resulting hybrid detector is presented as compatible with existing training-time OOD methods and is evaluated on datasets from the MultiOOD benchmark, where it reports average outperformance over prior state-of-the-art approaches.

Significance. If the reported average gains hold under scrutiny, the work usefully demonstrates that inference-time exploitation of modality-specific signals can improve OOD detection without retraining. It supplies a practical, plug-in enhancement rather than a new training objective, which could be adopted in multi-modal pipelines where robustness to distribution shift matters.

major comments (2)
  1. [§4] §4 (Experiments): the central claim of consistent improvement rests on average performance across the MultiOOD benchmark, yet the abstract and summary provide no per-dataset AUROC/FPR95 numbers, standard deviations, or statistical tests; without these, it is impossible to determine whether gains are uniform or driven by a subset of datasets.
  2. [§3] §3 (Method): the normalization of the feature-space score by multi-modal logits is asserted to preserve the OOD signal without new biases, but no ablation isolating this step or analysis of its effect on score distributions is referenced; this step is load-bearing for the hybrid detector's claimed advantage.
minor comments (2)
  1. [Abstract] The phrase 'interesting relationship' in the abstract and introduction should be replaced by a concise statement of the observed correlation (e.g., Pearson coefficient or qualitative pattern) to allow readers to assess its strength before the formal definition appears in §3.
  2. [Figures] Figure captions and axis labels in the experimental figures should explicitly state whether reported metrics are AUROC or FPR@95 and whether error bars reflect multiple runs or cross-validation folds.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below, providing clarifications from the manuscript and indicating where revisions will be made to strengthen the presentation.

read point-by-point responses

  1. Referee: [§4] §4 (Experiments): the central claim of consistent improvement rests on average performance across the MultiOOD benchmark, yet the abstract and summary provide no per-dataset AUROC/FPR95 numbers, standard deviations, or statistical tests; without these, it is impossible to determine whether gains are uniform or driven by a subset of datasets.

    Authors: The manuscript's Section 4 and associated tables report per-dataset AUROC and FPR95 values on the MultiOOD benchmark, with the average computed across them. The abstract emphasizes the average as the primary reported metric, which is standard for benchmark comparisons. We agree that explicit mention of consistency would improve clarity. In revision we will add a brief statement to the abstract noting that improvements hold on the majority of datasets and include standard deviations from repeated runs in the main results table. Statistical significance testing was not performed in the original submission but can be added if the editor deems it necessary. revision: partial

  2. Referee: [§3] §3 (Method): the normalization of the feature-space score by multi-modal logits is asserted to preserve the OOD signal without new biases, but no ablation isolating this step or analysis of its effect on score distributions is referenced; this step is load-bearing for the hybrid detector's claimed advantage.

    Authors: The normalization is introduced in Section 3 to scale the off-manifold feature score by the multi-modal logit magnitude, motivated by the observed relationship between uni- and multi-modal predictions. The text provides the mathematical justification but does not contain a dedicated ablation or distribution analysis for this component alone. We will add such an ablation (with and without normalization) together with score-distribution histograms in the revised manuscript to directly address this point. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an empirical observation of a relationship between multi-modal and uni-modal predictions, which is then used to construct a post-hoc hybrid OOD detector combined with feature-space scoring and logit normalization. No equations, derivations, fitted parameters renamed as predictions, or self-citations are shown that would reduce the detector score to its own inputs by construction. The approach is described as compatible with existing methods and validated externally on the MultiOOD benchmark, rendering the central claim self-contained without load-bearing circular reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no equations, training details, or modeling choices, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.1-grok · 5746 in / 1130 out tokens · 19229 ms · 2026-06-26T00:27:02.762874+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 3 linked inside Pith

  1. [1]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Ahmed, F., Courville, A.: Detecting semantic anomalies. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3154–3162 (2020)

    2020

  2. [2]

    Advances in Neural Information Processing Systems36, 38206–38230 (2023)

    Behpour, S., Doan, T.L., Li, X., He, W., Gou, L., Ren, L.: Gradorth: A simple yet efficient out-of-distribution detection with orthogonal projection of gradients. Advances in Neural Information Processing Systems36, 38206–38230 (2023)

    2023

  3. [3]

    arXiv preprint arXiv:1808.01340 (2018)

    Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. arXiv preprint arXiv:1808.01340 (2018)

    Pith/arXiv arXiv 2018

  4. [4]

    In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299–6308 (2017)

    2017

  5. [5]

    In: Proceedings of the IEEE international conference on computer vision

    Chéron, G., Laptev, I., Schmid, C.: P-cnn: Pose-based cnn features for action recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 3218–3226 (2015)

    2015

  6. [6]

    In: Proceedings of the European conference on computer vision (ECCV)

    Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Scaling egocentric vision: The epic-kitchens dataset. In: Proceedings of the European conference on computer vision (ECCV). pp. 720–736 (2018)

    2018

  7. [7]

    arXiv preprint arXiv:1802.04865 (2018)

    DeVries, T., Taylor, G.W.: Learning confidence for out-of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018)

    Pith/arXiv arXiv 2018

  8. [8]

    Advances in Neural Information Processing Systems36, 78674–78695 (2023)

    Dong,H.,Nejjar,I.,Sun,H.,Chatzi,E.,Fink,O.:Simmmdg:Asimpleandeffective framework for multi-modal domain generalization. Advances in Neural Information Processing Systems36, 78674–78695 (2023)

    2023

  9. [9]

    Advances in Neural Information Processing Sys- tems37, 129250–129278 (2024)

    Dong, H., Zhao, Y., Chatzi, E., Fink, O.: Multiood: Scaling out-of-distribution detection for multiple modalities. Advances in Neural Information Processing Sys- tems37, 129250–129278 (2024)

    2024

  10. [10]

    In: European Conference on Computer Vision

    Doorenbos, L., Sznitman, R., Márquez-Neila, P.: Data invariants to understand unsupervised out-of-distribution detection. In: European Conference on Computer Vision. pp. 133–150. Springer (2022)

    2022

  11. [11]

    arXiv preprint arXiv:2411.13619 (2024)

    Doorenbos, L., Sznitman, R., Márquez-Neila, P.: Non-linear outlier synthesis for out-of-distribution detection. arXiv preprint arXiv:2411.13619 (2024)

    arXiv 2024

  12. [12]

    Advances in Neural Information Processing Systems36(2024)

    Du, X., Sun, Y., Zhu, J., Li, Y.: Dream the impossible: Outlier imagination with diffusion models. Advances in Neural Information Processing Systems36(2024)

    2024

  13. [13]

    In: Proceedings of the International Conference on Learning Representations (2022)

    Du, X., Wang, Z., Cai, M., Li, Y.: Vos: Learning what you don’t know by vir- tual outlier synthesis. In: Proceedings of the International Conference on Learning Representations (2022)

    2022

  14. [14]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recog- nition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202–6211 (2019)

    2019

  15. [15]

    International Confer- ence on Machine Learning (2022) 16 L

    Hendrycks, D., Basart, S., Mazeika, M., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. International Confer- ence on Machine Learning (2022) 16 L. Doorenbos et al

    2022

  16. [16]

    Proceedings of International Conference on Learning Representations (2017)

    Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of- distribution examples in neural networks. Proceedings of International Conference on Learning Representations (2017)

    2017

  17. [17]

    International Conference on Learning Representations (2019)

    Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. International Conference on Learning Representations (2019)

    2019

  18. [18]

    Advances in Neural Information Processing Systems34, 677–689 (2021)

    Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distribu- tional shifts in the wild. Advances in Neural Information Processing Systems34, 677–689 (2021)

    2021

  19. [19]

    In: Forty-first International Conference on Machine Learning (2024)

    Huh, M., Cheung, B., Wang, T., Isola, P.: Position: The platonic representation hypothesis. In: Forty-first International Conference on Machine Learning (2024)

    2024

  20. [20]

    Kamoi, R., Kobayashi, K.: Why is the mahalanobis distance effective for anomaly detection? arXiv preprint arXiv:2003.00402 (2020)

    arXiv 2003

  21. [21]

    In: 2011 International conference on com- puter vision

    Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on com- puter vision. pp. 2556–2563. IEEE (2011)

    2011

  22. [22]

    In: 2011 IEEE international conference on robotics and automation

    Lai, K., Bo, L., Ren, X., Fox, D.: A large-scale hierarchical multi-view rgb-d object dataset. In: 2011 IEEE international conference on robotics and automation. pp. 1817–1824. IEEE (2011)

    2011

  23. [23]

    Journal of multivariate analysis88(2), 365–411 (2004)

    Ledoit, O., Wolf, M.: A well-conditioned estimator for large-dimensional covariance matrices. Journal of multivariate analysis88(2), 365–411 (2004)

    2004

  24. [24]

    International Conference on Learning Rep- resentations (2018)

    Lee, K., Lee, H., Lee, K., Shin, J.: Training confidence-calibrated classifiers for detecting out-of-distribution samples. International Conference on Learning Rep- resentations (2018)

    2018

  25. [25]

    Advances in neural information processing systems31(2018)

    Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out- of-distribution samples and adversarial attacks. Advances in neural information processing systems31(2018)

    2018

  26. [26]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Li, S., Gong, H., Dong, H., Yang, T., Tu, Z., Zhao, Y.: Dpu: Dynamic proto- type updating for multimodal out-of-distribution detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10193–10202 (2025)

    2025

  27. [27]

    Advances in Neural Information Processing Systems (2025)

    Liang, J., Hou, R., Hu, M., Chang, H., Shan, S., Chen, X.: Revisiting logit distri- butions for reliable out-of-distribution detection. Advances in Neural Information Processing Systems (2025)

    2025

  28. [28]

    In: Proceedings of the International Conference on Learning Representations (2018)

    Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: Proceedings of the International Conference on Learning Representations (2018)

    2018

  29. [29]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ling, Z., Chang, Y., Zhao, H., Zhao, X., Chow, K., Deng, S.: Cadref: Robust out-of-distribution detection via class-aware decoupled relative feature leveraging. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 4968–4977 (2025)

    2025

  30. [30]

    Advances in Neural Information Processing Systems (2025)

    Liu, M., Dong, H., Kelly, J., Fink, O., Trapp, M.: Extremely simple multimodal outlier synthesis for out-of-distribution detection and segmentation. Advances in Neural Information Processing Systems (2025)

    2025

  31. [31]

    Advances in Neural Information Processing Systems (2020)

    Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems (2020)

    2020

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of- distribution detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 23946–23955 (2023)

    2023

  33. [33]

    Internation Conference on Machine Learning (2025) Modality-Aware Out-of-Distribution Detection 17

    Mueller, M., Hein, M.: Mahalanobis++: Improving ood detection via feature nor- malization. Internation Conference on Machine Learning (2025) Modality-Aware Out-of-Distribution Detection 17

    2025

  34. [34]

    Neural Computing and Applications36(10), 5499–5513 (2024)

    Shaikh, M.B., Chai, D., Islam, S.M.S., Akhtar, N.: Multimodal fusion for audio- image and video action recognition. Neural Computing and Applications36(10), 5499–5513 (2024)

    2024

  35. [35]

    In: European Conference on Computer Vision

    Sharifi, S., Entesari, T., Safaei, B., Patel, V.M., Fazlyab, M.: Gradient-regularized out-of-distribution detection. In: European Conference on Computer Vision. pp. 459–478. Springer (2024)

    2024

  36. [36]

    arXiv preprint arXiv:1212.0402 (2012)

    Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)

    Pith/arXiv arXiv 2012

  37. [37]

    In: Proceedings of the International Conference on Machine Learning

    Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: Proceedings of the International Conference on Machine Learning. pp. 20827–20840 (2022)

    2022

  38. [38]

    IEEE transactions on pattern analysis and machine intelligence45(3), 3200–3225 (2022)

    Sun, Z., Ke, Q., Rahmani, H., Bennamoun, M., Wang, G., Liu, J.: Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence45(3), 3200–3225 (2022)

    2022

  39. [39]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Tang, K., Hou, C., Peng, W., Fang, X., Wu, Z., Nie, Y., Wang, W., Tian, Z.: Simpli- fication is all you need against out-of-distribution overconfidence. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5030–5040 (2025)

    2025

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4921–4930 (2022)

    2022

  41. [41]

    arXiv preprint arXiv:2007.05566 (2020)

    Winkens, J., Bunel, R., Roy, A.G., Stanforth, R., Natarajan, V., Ledsam, J.R., MacWilliams, P., Kohli, P., Karthikesalingam, A., Kohl, S., et al.: Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566 (2020)

    arXiv 2007

  42. [42]

    International Journal of Computer Vision132(12), 5635–5662 (2024)

    Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. International Journal of Computer Vision132(12), 5635–5662 (2024)

    2024

  43. [43]

    sword" sample happens in a gym court. ID classes such as “dribble

    Yang, Y., Xu, H.: Strengthen out-of-distribution detection capability with pro- gressive self-knowledge distillation. In: Forty-second International Conference on Machine Learning (2025) Modality-Aware Out-of-Distribution Detection 1 6 Full Experimental Details We provide more details on the dataset combinations used for the experiments. Alldatasetsfollow...

    2025