- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper focuses on open-world multi-view classification, simultaneously modeling the aleatoric uncertainty induced by noisy data and the epistemic uncertainty arising from unknown classes. The method represents each view with a Gaussian distribution, leverages KAN to collect evidence, and proposes UD-PCR to resolve cross-view conflicts. Closed-set and open-set experiments on six datasets demonstrate competitive overall performance.
Critical strengths
- The aleatoric uncertainty at the representation level and the epistemic uncertainty at the decision level are addressed within a unified framework.
- UD-PCR comes with a clear mathematical formulation and an analysis of its boundary properties. Experiments span six datasets and report OSCR, Open AUC, and low-FPR metrics, yielding a relatively comprehensive evaluation.
Critical weaknesses
- The open-set experiments are compared mainly against closed-set trustworthy multi-view methods, whereas open-world multi-view approaches discussed in the related work, such as OpenViewer and TOMC, are not included. Could the authors supplement this section with additional comparisons?
- The variance produced by the Gaussian encoder is directly interpreted as aleatoric uncertainty, yet no calibration metrics, correlation analysis with noise intensity, or quantitative perturbation experiments are provided to substantiate this interpretation.
Questions for authors
- How are known and unknown classes partitioned? Do the five runs adopt the same class split, and is an independent validation set used for model selection and threshold determination?
- Could the authors quantitatively verify the correlation between the Gaussian variance and input noise or prediction errors, and supplement an ablation study in which KAN is replaced solely by an MLP?
3 comments
Author Response to Reviewer gb1G — Part 2
[W2] The variance produced by the Gaussian encoder is directly interpreted as aleatoric uncertainty, yet no calibration metrics, correlation analysis with noise intensity, or quantitative perturbation experiments are provided to substantiate this interpretation.
Thank you for this valuable suggestion. We provide further clarification and additional experiments to substantiate the learned Gaussian variance as an indicator of view-specific aleatoric uncertainty.
(1) Aleatoric uncertainty modeling and noise correlation.
Our approach follows established probabilistic representation learning methods (Shi and Jain, ICCV 2019 [1]; Tang et al., ACL 2025 [2]) and recent distribution-based aleatoric uncertainty modeling (Tang et al., IJCAI 2025 [3]; Wang et al., CVPR 2026 [4]). Specifically, in our framework, each sample is mapped by encoders into a mean vector and a standard deviation vector, which jointly parameterize its latent Gaussian distribution. The mean captures view-specific semantic information, while the variance characterizes uncertainty associated with noisy or ambiguous observations, providing a principled basis for modeling view-specific aleatoric uncertainty.
Fig. 1. Analysis of Gaussian Variance and Adaptive Cross-View Fusion under Increasing Input Noise.
- Fig. 1(a): Correlation Between Gaussian Noise and Predicted Variance
- Fig. 1(b): View-Specific Predicted Variance under Increasing Noise
- Fig. 1(c): AU-DCF Adaptive View Weights under Increasing Noise
- Fig. 1(d): Open AUC Comparison Between AU-DCF and Uniform Fusion
To examine its relationship with input noise, we introduced varying levels of Gaussian noise into one view of the Scene dataset. As shown in Fig. 1(a–b), the change in predicted variance exhibits a positive correlation with noise intensity (Pearson ), while the average variance of the corrupted view increases with noise severity. Fig. 1(c–d) further shows that AU-DCF adaptively reduces the corrupted view's contribution and achieves better OSCR and Open AUC than uniform fusion.
(2) Quantitative perturbation experiments.
We further conducted experiments under varying Gaussian noise intensities (0–1.5), comparing EU²O with six representative TMVC baselines.
Fig. 2. Robustness Comparison under Varying Gaussian Noise Intensities on Three Datasets.
- Fig. 2(a): Closed-Set Classification Accuracy under Gaussian Noise
- Fig. 2(b): Open-Set Recognition Performance (OSCR) under Gaussian Noise
As shown in Fig. 2(a–b), EU²O generally demonstrates favorable robustness in both closed-set classification (Accuracy) and open-set recognition (OSCR and Open AUC) as noise intensity increases.
In summary, these results provide empirical support for the noise sensitivity of the learned Gaussian variance and the robustness of AU-DCF. We have incorporated these analyses into the revised manuscript.
References:
[1] Probabilistic Face Embeddings, ICCV 2019.
[2] Modeling Uncertainty in Composed Image Retrieval via Probabilistic Embeddings, ACL 2025.
[3] DFMU: Distribution-based Framework for Modeling Aleatoric Uncertainty in Multimodal Sentiment Analysis, IJCAI 2025.
[4] Uncertainty-Aware Modality Fusion for Unaligned RGB-T Salient Object Detection, CVPR 2026.
Author Response to Reviewer gb1G — Part 3
[Q1] How are known and unknown classes partitioned? Do the five runs adopt the same class split, and is an independent validation set used for model selection and threshold determination?
Thank you for the question. We clarify our data partitioning and evaluation protocol as follows.
(1) Known/unknown class partitioning.
Following the standard openness definition
we determine the number of known classes from the total number of classes and the target openness. The known classes are randomly selected, with 80% of their samples used for training and the remaining 20% for testing. All samples from unknown classes are exclusively reserved for testing and are never used during training.
(2) Five independent runs.
We perform five independent random partitions rather than using a fixed class split across all runs. Importantly, within each run, all competing methods share the same known/unknown class partition and training/test split to ensure fair comparisons. We report the mean and standard deviation across the five runs.
(3) Model selection and threshold determination.
We do not use an independent validation set for checkpoint selection or rejection-threshold tuning. Instead, each model is evaluated after completing its predefined training schedule. OSCR and Open AUC are computed by sweeping confidence thresholds on the held-out test predictions, while CCR@FPR is reported at predefined FPR levels (0.5% and 1%), rather than using a separately tuned deployment threshold.
We have clarified these implementation details in the revised manuscript.
[Q2] Could the authors quantitatively verify the correlation between the Gaussian variance and input noise or prediction errors, and supplement an ablation study in which KAN is replaced solely by an MLP?
Thank you for the suggestion. We have conducted additional experiments to address both concerns.
(1) Correlation between Gaussian variance and input noise.
We introduced Gaussian noise of varying intensities into View 3 of the Scene dataset. As shown in Fig. 1(a), the change in predicted variance exhibits a modest positive correlation with noise intensity (Pearson ). Moreover, the average predicted variance increases with noise severity (Fig. 1(b)), while AU-DCF adaptively reduces the corrupted view's fusion weight (Fig. 1(c)). Fig. 1(d) further demonstrates better Open AUC than uniform fusion under increasing noise.
These results provide empirical support for the noise sensitivity of the learned Gaussian variance, enabling AU-DCF to adaptively mitigate the influence of unreliable views and improve the robustness of cross-view fusion.
(2) KAN-to-MLP ablation study.
We further replaced only the KAN component with an MLP while keeping all other modules and experimental settings unchanged. The results on Scene are summarized below.
Table 2. Ablation study of the KAN-based evidential collector on Scene under different openness settings.
| Openness | Collector | [email protected]% ↑ | CCR@1% ↑ | OSCR ↑ | Open AUC ↑ |
|---|---|---|---|---|---|
| 0.2 | MLP | 5.28±4.32 | 7.70±6.21 | 61.73±4.99 | 66.62±5.67 |
| KAN | 5.54±4.03 | 8.42±5.73 | 62.08±4.24 | 67.62±5.26 | |
| 0.3 | MLP | 7.70±4.09 | 9.78±4.81 | 67.47±12.30 | 70.33±10.56 |
| KAN | 7.82±4.97 | 10.64±7.17 | 67.62±13.05 | 70.54±11.09 |
As shown in Table 2, KAN achieves higher performance across the evaluated metrics on Scene under different openness settings, providing additional improvements over the MLP-based collector.
We have incorporated these additional analyses and ablation results into the revised manuscript.
We hope our rebuttal clarifies and addresses all your concerns, and we also welcome any further constructive feedback you might provide to enhance our submission.
Author Response to Reviewer gb1G — Part 1
We sincerely thank you for your careful evaluation and constructive suggestions, which have helped us improve the clarity and completeness of our manuscript. We have carefully addressed your concerns through additional experiments, methodological clarifications, and corresponding revisions. Our detailed responses are provided below.
We sincerely thank you for this valuable suggestion.
(1) Comparison with OpenViewer.
Our initial comparisons primarily focus on existing TMVC methods, as EU²O aims to achieve reliable uncertainty estimation and trustworthy decision-making. OpenViewer [1] primarily addresses unknown-category recognition by generating pseudo-unknown samples to mitigate overconfident predictions.
To provide a more comprehensive evaluation, we further include OpenViewer as an additional open-set baseline. For fairness, we adapt OpenViewer to our unified experimental protocol (e.g., identical data partitions), while preserving its original architecture, loss functions, and core learning mechanisms.
Table 1. Comparison with OpenViewer under the unified experimental protocol.
As shown in Table 1, EU²O achieves higher mean OSCR and Open AUC across all six benchmarks under the unified protocol, further demonstrating its effectiveness in trusted open-set recognition.
(2) Discussion of TOMC [2].
Its official implementation was unavailable to us, and certain implementation details could not be reliably recovered from the paper alone. We discuss TOMC and its methodological differences in the Related Work section and further clarify these distinctions in the revision.
References:
[1] OpenViewer: Openness-Aware Multi-View Learning, ACM MM 2025.
[2] Trusted Open-World Multi-View Classification with Dynamic Opinion Aggregation, ACM MM 2025.