
DW
Daniel Wang
· 1 min read
ResearcharXiv cs.CV
Robust-GCD
arXiv:2611.18673v1 Announce Type: new
Abstract: Generalized Category Discovery (GCD) classifies an unlabeled pool that contains both known and novel classes, given labels for the known classes only. Existing GCD methods assume that every un- labeled image is clean, but unlabeled data collected in the wild is often degraded by blur, sensor noise and compression artifacts. We formalize the setting in which an unknown fraction of the unlabeled pool is corrupted by an unknown operator as Degraded Generalized Category Discovery (D-GCD). To tackle it, we propose Robust-GCD, a framework that estimates the corruption directly from the unlabeled pool and uses the estimate to build labeled anchors for known classes and pseudo-anchors for novel classes inside the corrupted region of feature space. A trust curriculum controls how much the unlabeled pool is trusted while the anchors take hold, and a paired alignment loss keeps a clean image and its corrupted twin close in feature space. Across five datasets and multiple corruption families, Robust-GCD recovers between 43% and 96% of the All accuracy that SimGCD loses under contamination, and more than 60% on four of the five datasets, with the largest gains on the corrupted images themselves.
Original source
This story was published by arXiv cs.CV and written by Daniel Wang. SyncAI.news shows a preview; the complete article is on the publisher's site.
Read the full story on arxiv.org


