Abstract
Vision systems trained in closed-world scenarios fail when presented with new environmental conditions, new data distributions, and novel classes at deployment time. How to move towards open-world learning is a long-standing research question. The existing solutions mainly focus on specific aspects of the problem (single domain Open-Set, multi-domain Closed-Set), or propose complex strategies which combine several losses and manually tuned hyperparameters. In this work, we tackle multi-source Open-Set domain adaptation by introducing HyMOS: a straightforward model that exploits the power of contrastive learning and the properties of its hyperspherical feature space to correctly predict known labels on the target, while rejecting samples belonging to any unknown class. HyMOS includes style transfer among the instance transformations of contrastive learning to get domain invariance while avoiding the risk of negative-transfer. A self-paced threshold is defined on the basis of the observed data distribution and updates online during training, allowing to handle the known-unknown separation. We validate our method over three challenging datasets. The obtained results show that HyMOS outperforms several competitors, defining the new state-of-the-art.
Introduction
Most computer vision models are trained in "closed-world" scenarios, assuming the classes seen during training are the only ones that will appear during deployment. In reality, models face two major hurdles: Domain Shift, where data comes from multiple sources with different styles (e.g., photos vs. sketches), and Category Shift (Open-Set), where the deployment environment contains novel objects belonging to classes the model has never seen before.
Existing solutions for Multi-source Open-Set Domain Adaptation (MS-OSDA) are often overly complex, relying on multiple loss functions and manually tuned hyperparameters. HyMOS (Hyperspherical Multi-source Open-Set) provides a streamlined framework that leverages the geometric properties of hyperspherical feature spaces to handle both domain alignment and unknown class rejection.
Methodology
The core of HyMOS is built on Supervised Contrastive Learning mapped onto a unit hypersphere. The approach consists of four key pillars:
Hyperspherical Embedding and Alignment
Features are projected onto a unit sphere, a geometry that encourages uniform distribution of class centers (prototypes) and provides a clear metric (cosine similarity) for classification. To align multiple sources, HyMOS uses a balanced sampling strategy in each mini-batch, ensuring every class and every source domain is represented. The contrastive loss then pulls samples of the same class together, regardless of their source domain.
Style-Transfer Augmentation
To make the model "blind" to domain-specific textures, HyMOS incorporates AdaIN (Adaptive Instance Normalization). During training, the style of a target image is randomly transferred onto a source image. This forces the encoder to focus on semantic content rather than domain-specific "style," effectively acting as a semantic filter.
Self-Paced Thresholding
Instead of a fixed threshold to reject "unknowns," HyMOS calculates a dynamic threshold based on class sparsity (inter-prototype distance) and class compactness (cluster tightness). This threshold updates online as the model learns, allowing for robust known-unknown separation.
Experiments and Results
The authors evaluated HyMOS on three standard benchmarks: Office31 (3 domains, 31 categories), Office-Home (4 domains, 65 categories), and DomainNet (6 domains, 345 categories). Performance was measured using the HOS metric (harmonic mean of accuracy on known classes and the unknown class).
HyMOS consistently outperformed existing state-of-the-art methods:
- DomainNet: HyMOS achieved an HOS of 59.3%, significantly higher than the previous best competitor, ROS (48.5%).
- Office-Home: It reached 67.2% HOS, outperforming the multi-source baseline MOSDANET (59.1%).
Ablation studies confirmed the necessity of each component: removing style transfer caused a 3.6% drop in performance, while removing balanced source sampling led to a 2.6% drop.
Conclusion
HyMOS demonstrates that a straightforward approach grounded in feature space geometry can outperform complex, multi-loss systems for MS-OSDA. By treating style as an augmentation and leveraging the "no man's land" of hyperspheres for rejection, it provides a robust solution for open-world learning. Future work includes reducing dependency on external style models and further refining self-training to mitigate model drift.
Summary updated