Dual-Stream Adapters for Anomaly Segmentation

Shyam Nandan Rai, Massimiliano Mancini, Marco Ciccone, Iuliana Georgescu, Zeynep Akata, Barbara Caputo, Carlo Masone

Keywords: Adapters, Anomaly Segmentation

Abstract: Anomaly segmentation aims to identify pixels of objects not present during the model’s training. Recent approaches address this task using mask-based architectures, but these methods have high training costs due to the large transformer backbones involved. While vision adapters can help reduce training costs, they are not specialized for this task, leading to inferior performance. In this work, we propose Dual-Stream Adapters (DSA), a vision adapter tailored for anomaly segmentation. DSA extracts both in-distribution and out-of-distribution features via (i) an anomaly prior module that produces separate initial embeddings for the two streams; and (ii) a dual-stream feature refinement that implicitly guides the separation of in-distribution from out-of-distribution features. We train DSA using a novel hyperbolic loss function that provides supervised guidance for differentiating in-distribution and out-of-distribution features. Experiments on various benchmarks show that dual-stream adapters achieve the best results while reducing training parameters by 38\% w.r.t. the previous state-of-the-art.

Primary Area: applications to robotics, autonomy, planning

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.

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: 7865