CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Lafon, Marc, Hakim, Gustavo Adolfo Vargas, Rambour, Clément, Desrosier, Christian, Thome, Nicolas
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916956922183680
author Lafon, Marc
Hakim, Gustavo Adolfo Vargas
Rambour, Clément
Desrosier, Christian
Thome, Nicolas
author_facet Lafon, Marc
Hakim, Gustavo Adolfo Vargas
Rambour, Clément
Desrosier, Christian
Thome, Nicolas
contents Vision-language models (VLMs) like CLIP exhibit strong zero-shot capabilities but often fail to generalize under distribution shifts. Test-time adaptation (TTA) allows models to update at inference time without labeled data, typically via entropy minimization. However, this objective is fundamentally misaligned with the contrastive image-text training of VLMs, limiting adaptation performance and introducing failure modes such as pseudo-label drift and class collapse. We propose CLIPTTA, a new gradient-based TTA method for vision-language models that leverages a soft contrastive loss aligned with CLIP's pre-training objective. We provide a theoretical analysis of CLIPTTA's gradients, showing how its batch-aware design mitigates the risk of collapse. We further extend CLIPTTA to the open-set setting, where both in-distribution (ID) and out-of-distribution (OOD) samples are encountered, using an Outlier Contrastive Exposure (OCE) loss to improve OOD detection. Evaluated on 75 datasets spanning diverse distribution shifts, CLIPTTA consistently outperforms entropy-based objectives and is highly competitive with state-of-the-art TTA methods, outperforming them on a large number of datasets and exhibiting more stable performance across diverse shifts.
format Preprint
id arxiv_https___arxiv_org_abs_2507_14312
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation
Lafon, Marc
Hakim, Gustavo Adolfo Vargas
Rambour, Clément
Desrosier, Christian
Thome, Nicolas
Computer Vision and Pattern Recognition
Vision-language models (VLMs) like CLIP exhibit strong zero-shot capabilities but often fail to generalize under distribution shifts. Test-time adaptation (TTA) allows models to update at inference time without labeled data, typically via entropy minimization. However, this objective is fundamentally misaligned with the contrastive image-text training of VLMs, limiting adaptation performance and introducing failure modes such as pseudo-label drift and class collapse. We propose CLIPTTA, a new gradient-based TTA method for vision-language models that leverages a soft contrastive loss aligned with CLIP's pre-training objective. We provide a theoretical analysis of CLIPTTA's gradients, showing how its batch-aware design mitigates the risk of collapse. We further extend CLIPTTA to the open-set setting, where both in-distribution (ID) and out-of-distribution (OOD) samples are encountered, using an Outlier Contrastive Exposure (OCE) loss to improve OOD detection. Evaluated on 75 datasets spanning diverse distribution shifts, CLIPTTA consistently outperforms entropy-based objectives and is highly competitive with state-of-the-art TTA methods, outperforming them on a large number of datasets and exhibiting more stable performance across diverse shifts.
title CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2507.14312