Domain-Held-Out Validation and Temperature Calibration: Stabilizing Rejection Performance of CLIP Adapters for Open-Set Domain Generalization

Abstract Open-set recognition (OSR) systems must correctly classify known classes while rejecting unknown-class inputs. In practical image-processing pipelines, OSR is frequently deployed under domain shift (e.g., photo vs.\sketch), where the \emph{operating point} of a rejection module can become unstable: the scale and shape of logit/score distributions drift across domains, so a threshold tuned on one domain may yield a higher false-alarm rate or missed rejections on another. Motivated by this reliability issue, we study CLIP-based open-set domain generalization (OSDG) on the PACS benchmark and use a lightweight CLIP adapter trained on source-domain labeled known classes to improve closed-set discrimination. We then propose \textbf{DHV-DS} (\emph{Domain-Held-out Validation with Decoupled Scoring}), a minimal calibration pipeline that explicitly accounts for domain shift: (1) \emph{Domain-held-out validation} holds out an entire source domain as a validation domain that is never used for training; (2) \emph{Decoupled scoring with automatic selection} keeps adapter logits for classification, while allowing the rejection score to be computed from zero-shot logits, adapter logits, or their mixture, selected automatically on the held-out domain; (3) \emph{Joint temperature and threshold calibration} tunes temperature and chooses a rejection threshold on the held-out domain to better transfer the desired operating point. On PACS (macro-averaged over four target domains, 10 unknown-class splits, and two random seeds), DHV-DS reduces FPR@95 from 0.1148 to 0.1004 and improves AUROC from 0.9606 to 0.9677, achieving the lowest mean FPR@95 on all four target domains. Paired comparisons show consistent gains across unknown partitions and seeds. DHV-DS requires no extra data and adds negligible overhead, making it a practical default procedure for stabilizing cross-domain OSR rejection in CLIP-adapter-based systems.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-8631400/v1
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Domain-Held-Out Validation and Temperature Calibration: Stabilizing Rejection Performance of CLIP Adapters for Open-Set Domain Generalization

Ke Jang, Taoze Liu, Hongyi Yang, Jiegang Liu et al.
3 links
Research Square
Domain Adaptation and Few-Shot Learning
preprint

Domain-Held-Out Validation and Temperature Calibration: Stabilizing Rejection Performance of CLIP Adapters for Open-Set Domain Generalization

Ke Jang, Taoze Liu, Hongyi Yang, Jiegang Liu, Cheng Fang, Jin Zhang, Zhanghong Wang, Shuang Li
preprint en

Abstract

Abstract Open-set recognition (OSR) systems must correctly classify known classes while rejecting unknown-class inputs. In practical image-processing pipelines, OSR is frequently deployed under domain shift (e.g., photo vs.\sketch), where the \emph{operating point} of a rejection module can become unstable: the scale and shape of logit/score distributions drift across domains, so a threshold tuned on one domain may yield a higher false-alarm rate or missed rejections on another. Motivated by this reliability issue, we study CLIP-based open-set domain generalization (OSDG) on the PACS benchmark and use a lightweight CLIP adapter trained on source-domain labeled known classes to improve closed-set discrimination. We then propose \textbf{DHV-DS} (\emph{Domain-Held-out Validation with Decoupled Scoring}), a minimal calibration pipeline that explicitly accounts for domain shift: (1) \emph{Domain-held-out validation} holds out an entire source domain as a validation domain that is never used for training; (2) \emph{Decoupled scoring with automatic selection} keeps adapter logits for classification, while allowing the rejection score to be computed from zero-shot logits, adapter logits, or their mixture, selected automatically on the held-out domain; (3) \emph{Joint temperature and threshold calibration} tunes temperature and chooses a rejection threshold on the held-out domain to better transfer the desired operating point. On PACS (macro-averaged over four target domains, 10 unknown-class splits, and two random seeds), DHV-DS reduces FPR@95 from 0.1148 to 0.1004 and improves AUROC from 0.9606 to 0.9677, achieving the lowest mean FPR@95 on all four target domains. Paired comparisons show consistent gains across unknown partitions and seeds. DHV-DS requires no extra data and adds negligible overhead, making it a practical default procedure for stabilizing cross-domain OSR rejection in CLIP-adapter-based systems.

Research Square
Guizhou Minzu University (CN)
Guizhou Minzu University
Domain Adaptation and Few-Shot Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon