MoDA: Mixture of Domain Adapters for Parameter-efficient Generalizable Person Re-Identification
Keywords: Generalizable Person Re-Identification, Domain Generalization, Parameter-efficient Fine-tuning
Abstract: Domain Generalizable ReID task has garnered much attention in recent years, as a more challenging task but more closely aligned with practical applications. Mixture-of-experts (MoE) based methods has been studied for DG ReID to exploit the discrepancies and inherent correlations between diverse domains. However, most of DG ReID methods, including MoE-based methods, have to full fine-tune the large amount of parameters of backbones, classifier heads and experts. And in the set of DG ReID, the number of person IDs is particularly large which results in that parameters of classifier heads increases sharply. And make it difficult for MoE-based method to scale up to larger vision models. For this motivation, we propose a novel MoE-based DG ReID method, named mixture of do- main adapters (MoDA), to mitigate the issues men- tioned above. We apply adapter and CLIP to DG ReID in a parameter-efficient way. Extensive experiments verify that MoDA achieves competitive end even better results with state-of-the-art methods with much fewer tunable parameters.
Primary Area: representation learning for computer vision, audio, language, and other modalities
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/2024/AuthorGuide.
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: 220
Loading