Research map / Label-efficient and robust learning / OOD and anomaly detection

Self-Calibrated Tuning of Vision-Language Models for Out-of-Distribution Detection

NeurIPS 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Spurious-aware OOD detection · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Self-Calibrated Tuning (SCT), which adds uncertainty-based modulating factors to the LoCoOp prompt-tuning objective so that OOD regularization from imperfect ID-derived outlier features is calibrated per sample, improving few-shot CLIP OOD detection on ImageNet-1k.

Problem

Prompt-tuning-based OOD detection with CLIP regularizes prompts using surrogate OOD features mined from ID images, but imperfect foreground-background decomposition makes many of these features spurious, especially for high-uncertainty ID samples, hurting OOD detection.

Approach

Multiply the classification cross-entropy by (1 - p(y|x)) and the OOD regularization term by p(y|x), so per-sample uncertainty redirects optimization: low-confidence samples emphasize classification while confident samples emphasize OOD regularization.

Key results

• On ImageNet-1k with 16-shot training, SCT reaches average FPR95 26.47 and AUROC 93.37, versus LoCoOp's 29.47 and 93.10.
• On ImageNet-1k with 1-shot training, SCT reaches average FPR95 31.09 versus LoCoOp's 34.94.
• Adding SCT to IDLike and LSN improves their average FPR95 in the compatibility experiments (e.g., LSN 39.58 to 36.85 at 16-shot).

Measured effects reported in the paper

Surprising findings

• Setting the regularization weight lambda=1 may degrade OOD detection, suggesting classification should get more weight.
• Under 1-shot, SCT's AUROC gain over LoCoOp is small (92.04 vs 91.75), and the paper notes AUROC improvements are less notable than FPR95.

Limitations

• No theoretical justification for the choice of linear modulating functions.
• Few-shot results are sensitive to the quality of the limited ID training data.
• Gains on AUROC are smaller than on FPR95.

Experimental setup

Datasets: ImageNet-1k iNaturalist SUN Places365 Textures CIFAR-10 CIFAR-100

Benchmarks: ImageNet-1k OOD benchmarks (iNaturalist, SUN, Places365, Textures) ImageNet-10/20/100 hard OOD CIFAR-10/CIFAR-100 OOD

Metrics: FPR95 AUROC ID-ACC ECE

Baselines: MCM GL-MCM MSP ODIN Energy ReAct MaxLogit CoOp LoCoOp IDLike NegPrompt LSN

Builds on: CoOp LoCoOp MCM GL-MCM

Closest papers by meaning

• Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection (ICLR 2025 · similarity 0.87)
• Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)
• ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection (CVPR 2024 · similarity 0.87)
• Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.87)
• Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · similarity 0.87)
• SeTAR: Out-of-Distribution Detection with Selective Low-Rank Approximation (NeurIPS 2024 · similarity 0.87)