The field of Domain Generalization (DG) is vital for addressing domain shifts in critical real-world applications such as healthcare, autonomous driving, and environmental monitoring, where the variability of conditions is both frequent and unpredictable. Despite considerable progress, much of the existing research in DG remains limited to relatively simple scenarios—often confined to single tasks, single modalities, and straightforward domain shifts within closed-world environments. This narrow scope overlooks the complexities inherent in real deployments, which frequently involve multi-modal inputs, a variety of tasks, innovative neural architectures, extreme environmental conditions, and severe resource constraints. Our workshop, Domain Generalization: Evolution, Breakthroughs, and Future Horizons, aims to map the trajectory of DG from its foundational breakthroughs to the next frontiers. We will explore how DG can evolve to handle the nuanced challenges of modern applications, fostering a dialogue that encourages innovative research approaches and cross-disciplinary collaborations. By providing a platform for discussing both current advancements and future directions, our workshop seeks to catalyze new ideas and methodologies that will significantly extend the robustness and applicability of DG to more complex and diverse real-world conditions. 

Potential Topics:

• Expanding Domain Generalization Beyond Classification and Images: Exploring domain generalization in more complex tasks like object detection and visual question answering (VQA), extending beyond static images to video, and from 2D to 3D vision tasks such as monocular depth estimation.

• Tackling Extreme Shifts and Long-tail Events: Addressing the challenges posed by extreme distribution shifts and long-tail events, which are difficult to represent in large datasets, aiming to enhance robustness in unpredictable scenarios.

• Integrating Domain Generalization with Model Calibration: Proposing a combined
approach to domain generalization and model calibration to improve robustness and reliability across varied real-world applications.

• Generalizing Multimodal Models Across Domains: Investigating how multimodal models, especially vision-language models, perform across different domains and assessing the impact of domain shifts on one or more modalities.

• Domain Generalization in Open-World Settings: Developing methods that enable models to adaptively recognize and classify new, unseen classes in varying environments.

• Enhancing Multitask Model Generalization: Improving the domain generalization capabilities of multitask models, balancing sensitivity to domain shifts across different tasks.

• Domain Generalization Beyond Natural Images and Supervised Settings: Exploring domain generalization in specialized applications such as remote sensing and medical imaging, where domain shifts are prevalent and labeled data is scarce.

Accepted papers will be presented at the poster session, some as orals and one paper will be awarded as the best paper.

Submission Guidelines:

• We accept submissions of max 8 pages (excluding references). We encourage authors to submit 4 page work as well.

• We accept dual submissions to CVPR 2025 and Domain Generalization: Evolution , Breakthroughs and Future Horizon workshop.

• Submitted manuscript should follow the CVPR 2025 paper template.

• Submissions will be rejected without review if they:

  1. Contain more than 8 pages (excluding references).

  2. Violate the double-blind policy.

  3. Violate the dual-submission policy for papers with more than 4 pages excluding references.

• The accepted papers will be in the main conference workshops proceedings if the authors agree (this option is valid only for full-length papers not published at CVPR 2025).

• Short papers will be linked at the workshop webpage but will not be in the main conference proceedings.

• Papers will be peer reviewed under double-blind policy, and must be submitted online through the CMT submission system.