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

Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution

NeurIPS 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Provable and feature-based OOD detection · 0 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 ImOOD, a statistical framework and unified training-time regularization that corrects a class-aware bias in OOD detectors trained on long-tailed ID data, improving OOD detection on CIFAR10/100-LT and ImageNet-LT.

Problem

OOD detection performance degrades substantially when the in-distribution training data is class-imbalanced, with tail-class ID samples flagged as OOD and OOD samples predicted as head classes.

Approach

A probabilistic derivation shows a class-aware bias term beta(x) between balanced and imbalanced OOD posteriors; a learnable class-aware factor gamma is used to calibrate the OOD logit during training via a BCE loss on calibrated logits plus a constraint loss, with no extra cost at inference.

Key results

• On CIFAR10-LT, COCL + Ours reaches 93.55 AUROC and 28.52 FPR95, versus 93.28 AUROC and 30.88 FPR95 for COCL.
• On ImageNet-LT, PASCL + Ours improves AUROC from 68.00 to 74.69 and reduces FPR95 from 87.53 to 74.37; COCL + Ours reaches 75.84 AUROC.
• On CIFAR10-LT, adding the regularization to BinDisc raises AUROC from 90.06 to 92.23 and lowers FPR95 from 33.39 to 29.95.

Measured effects reported in the paper

Surprising findings

• On CIFAR100-LT, the PASCL + Ours variant has lower AUPR than the COCL baseline and COCL + Ours has higher FPR95 than COCL alone, so gains are not uniform across metrics.
• Applying a fixed gamma=1 bias correction at inference only slightly improves BinDisc on CIFAR10-LT (AUROC 90.06 to 90.34), indicating a learned input- and class-dependent gamma is needed.

Limitations

• Relies on auxiliary OOD training samples (e.g., TinyImages80M, ImageNet-Extra), so unforeseen OOD samples at test time may challenge the learned boundary.
• Imbalance problem is not fully solved; tail-class scarcity still affects data-driven learning, and data-level re-balancing is suggested.
• Inference-time variant without training gives only marginal gains, and the input-dependent gamma estimate needs training.

Experimental setup

Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT TinyImages80M ImageNet-Extra ImageNet-1k

Benchmarks: CIFAR10-LT CIFAR100-LT ImageNet-LT SC-OOD ClassPrior ImageNet-LT-a8

Metrics: AUROC AUPR FPR95 ACC

Baselines: MSP OECC EnergyOE OE PASCL OpenSampling ClassPrior BalEnergy EAT COCL

Builds on: PASCL COCL Outlier Exposure Energy-based OOD detection BinDisc

Closest papers by meaning

• Long-Tailed Out-of-Distribution Detection via Normalized Outlier Distribution Adaptation (NeurIPS 2024 · similarity 0.91)
• EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.91)
• Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.90)
• Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.89)
• Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.89)
• Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.89)