RouGE: Learning Gated Experts for Segment Anything in the Wild
Keywords: PEFT，real-world degradation， segment anything model， robustness enhancement
TL;DR: We propose a novel plug-in module called RouGE, using paramter-efficient method to directly enhance the robustness of pre-trained SAM-based model against real-world degradation, empowering an All-in-One segmentation model.
Abstract: Segment anything model (SAM) and its variants have recently shown promising performance as foundation models. However, existing SAM-based models can only handle scenarios seen during training, and usually suffer unstable performance when transferring to real-world unseen data, such as low-light, rainy or blurred images, which is crucial for applications such as autopilot. Therefore, adapting SAM-based models for real-world degradation while not impairing its original ability remains an open challenge. In this work, we propose a novel gated Mixture-of-Experts (MoE) structure, called RouGE, to improve the robustness of SAM-based models. Specifically, RouGE uses multiple lightweight probability gates to decompose complex real-world image conditions and judge whether the feature needs to be adjusted as well as to what extent the adjustment needs to be done, then handle them differently with a set of low-rank experts. During the inference stage, RouGE processes input images in a completely blind manner thus improving the model's performance in real-world scenarios. Extensive experiments demonstrate that RouGE consistently achieves state-of-the-art results on both degraded and clean images compared with other methods while tuning only 1.5% of parameters.
Supplementary Material: zip
Primary Area: applications to 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/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: 5503
Loading