Existing Source-free Video Domain Adaptation (SFVDA) aims to learn a target video model for an unlabeled target domain by transferring knowledge from a labeled source domain using a single pre-trained source video model. In this paper, we explore a new SFVDA setting where multiple source domains exist, each offering a library of source models with different architectures. This setting offers both opportunities and challenges: while the presence of multiple source models enriches the pool of transferable knowledge, it also increases the risk of negative transfer due to inappropriate source knowledge. To tackle these challenges, we introduce the Multiple-Source-Video-Model Aggregation Framework (MSVMA), comprising two key modules. The first module, termed Multi-level Instance Transferability Calibration (MITC), enhances existing uncertainty-based transferability estimation metrics by incorporating scale information from both group and dataset levels. This integration facilitates accurate transferability estimation at the instance level across diverse models. The second module, termed Instance-level Multi Video Model Aggregation (IMVMA), leverages the calculated instance-level transferability to guide a path generation network. This network produces instance-specific weights for unsupervised aggregation of source models. Empirical results from three video domain adaptation datasets demonstrate the state-of-the-art performance of our MSVMA framework.
Keywords: Domain Adaptation, Multi-Task Learning
Abstract:
Supplementary Material: zip
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3045
Loading