Data-Centric Learning for Visual and Multimodal Foundation Models
Half-Day Workshop @ WACV 2027
The next generation of visual and multimodal foundation models will be shaped not only by scale, but by the quality, composition, alignment, and evolution of their data. Significant challenges remain in ethically constructing high-quality datasets, assessing data quality, aligning information across modalities, and continually refining data as models and deployment requirements evolve.
Data4VFM brings together researchers and practitioners developing data-centric approaches for visual and multimodal foundation models — with a focus on ethical dataset construction, data quality assessment, multimodal alignment, continual dataset improvement, synthetic data generation, and dataset construction for efficient pre-training, post-training, alignment, and adaptation of these models.
Note:
Submission details and deadlines are maintained through the OpenReview workshop page.
We welcome submissions across the following areas (non-exhaustive).
Image/video dataset construction, auditing, benchmark design, data provenance, licensing, privacy, and governance.
Coreset methods, subset selection, dataset distillation, filtering, deduplication, data valuation, efficient dataset scaling strategies.
Active learning, human-in-the-loop annotation, scaling laws, long-tail discovery, failure analysis, data attribution, robustness evaluation, alignment training.
Image-text and video-text datasets, 3D and point cloud datasets, datasets for multimodal pre-training/post-training/alignment, temporal data curation, cross-modal supervision, video-language learning.
Diffusion-generated images and videos, simulation environments, data augmentation, synthetic data quality assessment, hybrid real-synthetic datasets.
Dataset exploration tools, large-scale data pipelines, human-in-the-loop workflows, production deployments, autonomous driving use cases, and industrial case studies.
We welcome short papers (4 pages + references and appendices), including original research, ongoing work, and industrial case studies related to ethical dataset construction, data selection & pruning, data-centric learning & evaluation, multimodal & video & 3D foundation datasets, synthetic data, and systems & industrial practice (including autonomous driving use cases). Accepted papers will be presented as posters, with select papers invited for contributed talks.
Required format
All papers must use the official WACV 2027 author kit and paper template.
Questions? Contact data4vfm_wacv2027_workshop@googlegroups.com
Interested in reviewing? Sign up as a reviewer
All listed submission times are 11:59 AM UTC.
Submission portal opens via OpenReview.
Deadline for paper submissions.
Authors are notified of accept/reject decisions.
Final camera-ready versions of accepted papers are due.
Half-day workshop at WACV 2027, Disney Springs, Buena Vista, FL.
The main paper is limited to 4 pages. References and appendices may be any length and do not count toward the 4-page limit.
Accepted papers are non-archival and will not be published in the WACV proceedings.
Yes. Appendices can be included in the same PDF as the main paper and may contain any number of pages.
We welcome original research, work in progress, and industrial case studies related to data-centric learning for visual and multimodal foundation models.
Yes. Submissions must follow WACV's double-blind review guidelines and use the official WACV 2027 template.
Accepted papers will be presented as posters, with select papers invited for contributed talks.
