Abstract
Existing multi-view learning models struggle in open-set scenarios due to their implicit assumption of class completeness. Moreover, static view-induced biases, which arise from spurious view-label associations formed during training, further degrade their ability to recognize unknown categories. In this paper, we propose a multi-view open-set learning framework via ambiguity uncertainty calibration and view-wise debiasing. To simulate ambiguous samples, we design O-Mix, a novel synthesis strategy to generate virtual samples with calibrated open-set ambiguity uncertainty. These samples are further processed by an auxiliary ambiguity perception network that captures atypical patterns for improved open-set adaptation. Furthermore, we incorporate an HSIC-based contrastive debiasing module that enforces independence between view-specific ambiguous and view-consistent representations, encouraging the model to learn generalizable features. Extensive experiments on diverse multi-view benchmarks demonstrate that the proposed framework consistently enhances unknown-class recognition while preserving strong closed-set performance.
Introduction
Most traditional multi-view learning (MVL) models operate under a "closed-world" assumption, expecting that every object encountered during testing belongs to a category seen during training. In real-world applications, however, AI systems frequently encounter "open-set" scenarios containing previously unseen categories.
Two primary hurdles hinder current MVL models in these environments:
- Overconfidence in Unknowns: Models tend to force-rank unknown samples into known categories with high confidence.
- Static View-Induced Bias: Models often learn spurious associations (e.g., associating the color green with the category "dog" because training images of dogs were taken on grass). While these associations may hold for known classes, they lead to failure when unknown classes share these biased features.
The MOCD Framework
To address these challenges, we introduce MOCD (Multi-view Open-set Calibration and Debiasing), a framework designed to explicitly model uncertainty and decouple view-specific biases from generalizable features.
Core Components
- O-Mix (Open-set Mixup): Unlike standard Mixup, O-Mix utilizes Dempster-Shafer (DS) Theory to calibrate uncertainty. It synthesizes virtual training samples by mathematically assigning "mass" to both "ambiguous" and "unknown" states, effectively teaching the model to recognize the boundaries of its own knowledge.
- Ambiguity Perception Network (APN): This auxiliary branch processes synthesized O-Mix samples to capture atypical patterns that do not align with the prototypes of known classes, enhancing open-set adaptation.
- Contrastive Debiasing: Using the Hilbert-Schmidt Independence Criterion (HSIC), the framework enforces mathematical independence between view-consistent representations and view-specific ambiguous features. This forces the model to ignore spurious, view-induced biases and focus on robust, cross-view features.
Experimental Evaluation
The framework was evaluated across six diverse benchmarks, including BBCNews, Caltech20, NUSWIDE-OBJ, Hdigit, Iaprtc12, and VGGFace2. Experiments were conducted with an "Openness" setting of 10%, meaning 10% of test-time classes were entirely new to the model.
Key Results
MOCD consistently outperformed state-of-the-art baselines, particularly in the Correct Classification Rate (CCR) at low False Positive Rates (FPR). For instance, on the VGGFace2 dataset at a strict 1% FPR, MOCD achieved a 27.40% CCR, nearly doubling the performance of the next best baseline, TUNED (15.36%).
Crucially, MOCD maintains high accuracy on known classes while excelling at unknown-class rejection. Qualitative analysis via t-SNE visualizations confirms that MOCD produces more distinct clusters for known categories and effectively isolates unknown samples into a separate, scattered space.
Conclusion and Future Work
The MOCD framework demonstrates that treating uncertainty as a primary objective and explicitly debiasing view-specific features significantly improves multi-view open-set performance. While the framework introduces some computational overhead due to its multi-branch architecture, its "plug-and-play" potential allows it to enhance existing MVL models. Future research will explore adapting this framework for unsupervised or semi-supervised settings where labeled data is limited.
Summary updated