Research map / Label-efficient and robust learning / OOD and anomaly detection
Image Background Serves as Good Proxy for Out-of-distribution Data
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 a probabilistic framework that factorizes posterior into ID classification and OOD factor, and SSOD, an end-to-end OOD-data-free method that samples natural OOD supervision from background patches of ID feature maps.
Problem
Out-of-distribution detection for classifiers trained on closed image sets, where existing methods rely on frozen pre-trained features with limited OOD supervision or on costly synthetic OOD generation.
Approach
Use the local property of convolutional feature maps: patches with low ground-truth-class confidence are treated as natural OOD samples and high-confidence patches as ID samples, training an OOD head jointly with classification without explicit OOD data or generation.
Key results
- On ImageNet, SSOD averages 31.89% FPR95 and 92.26% AUROC across four OOD datasets, -6.28% FPR95 and +0.77% AUROC over the next best method.
- On CIFAR-10 (ResNet-18), SSOD averages 5.19% FPR95 and 98.94% AUROC, improving FPR95 by 19.01% and AUROC by 3.04% over the best prior method.
- On hard OOD, SSOD achieves top-ranked performance on OpenImage-O but is 4.38% worse in FPR95 than ViM on ImageNet-O's average comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SSOD self-supervised background-patch OOD supervision | ResNet-50 · ImageNet OOD average (iNaturalist, SUN, Places, Texture) | FPR95: -6.28% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-50 · ImageNet OOD average (iNaturalist, SUN, Places, Texture) | AUROC: +0.77% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-18 · CIFAR-10 OOD average | FPR95: -19.01% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-18 · CIFAR-10 OOD average | AUROC: +3.04% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-18 · Places (CIFAR-10 ID) | FPR95: -30.87% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-18 · Texture (CIFAR-10 ID) | FPR95: -21.30% | Helped |
| Add SSOD self-supervised background-patch OOD supervision | ResNet-50 · Texture (ImageNet ID) | FPR95: 33.76% | Hurt |
| Increase the OOD-head loss weight α (ablation) | SSOD with ResNet-18 · CIFAR-10 ID, Places OOD | FPR95 | Helped |
| Increase the OOD-head loss weight α (ablation) | SSOD with ResNet-18 · CIFAR-10 ID, Places OOD | ID ACC | Hurt |
| Use loss-wise balance (LWB) for ID/OOD feature imbalance (ablation) | SSOD · CIFAR-10 ID, Places OOD | OOD detection | Mixed |
Surprising findings
- SSOD performs worse than KNN on Texture (about 33.76% higher FPR95) because Texture images contain object-like symbols overlapping ImageNet classes.
- SSOD is weaker than MaDist on iSUN in CIFAR-10 experiments.
Limitations
- Performance limited when OOD data shares domain or object symbols with ID training images (e.g., Texture).
- Relies on diversity of background in training images to supply OOD supervision.
- Hard OOD on ImageNet-O (adversarial images) remains weaker than best prior methods.
Experimental setup
Datasets: ImageNet CIFAR-10
Benchmarks: ImageNet OOD (iNaturalist, SUN, Places, Texture) CIFAR-10 OOD (SVHN, LSUN, iSUN, Texture, Places) ImageNet-O OpenImage-O
Metrics: FPR95 AUROC ID ACC
Baselines: MSP MaDist ODIN KLM MaxLogit GODIN CSI MOS Energy KNN MCM Fort
Builds on: Energy KNN ReAct VOS MOS
Closest papers by meaning
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.92)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.92)
- Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.90)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.90)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.90)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.