VUOD: A Versatile Unsupervised Outlier Detection Framework for Natural, Industrial, Medical Images and Beyond

Accepted by ECCV 2026

Zhonghang Liu¹ , Siyuan Chen¹ , Jingwen Yu² , Changshuo Wang³ , Kunyang Li^4,‡ , Jiangbo Lu^1,✉

¹SmartMore Corporation, China ²The Hong Kong University of Science and Technology, China ³University College London, UK ⁴Rice University, USA

^‡ Work done during an internship at SmartMore.

^✉ Corresponding author: jiangbo.lu@gmail.com

Paper Code

Teaser

Left plot: VUOD significantly outperforms existing SOTAs on UOD benchmark datasets and real-world applications. More importantly, VUOD can be seamlessly integrated into various downstream visual tasks, such as image classification (middle plot) and 3D reconstruction with VGGT (right plot).

Abstract

Unsupervised outlier detection that automatically identifies whether visual systems involve anomalous images is a significant research topic. However, most approaches are limited to natural images because they rely on frozen and pre-trained feature extractors. So that their performance cannot be maintained in practical scenarios, especially for industrial inspection and medical imaging. In this work, we first revisit this task and then introduce a versatile unsupervised outlier detection framework to enrich the application domains. The core idea of this framework is to improve feature discriminativeness via exploiting intrinsic distribution priors. Evaluated on 3 domains and 14 benchmark datasets, our proposed solution achieves state-of-the-art performance and significantly outperforms existing methods. More importantly, we show its plug-and-play property that can be integrated into diverse visual applications to improve their robustness, such as image classification and 3D reconstruction.

Methodology

To apply UOD approaches into diverse domains, the core idea is to enhance the discriminativeness of feature representation via pseudo labeling (self-supervision generation) , i.e., splitting both inlier features and outlier features from features of the target dataset, and thereby leveraging contrastive learning strategy to enlarge the similarity within inliers and the difference between inliers and outliers in the enhanced feature space.

Local Separability

We adopt Affinity Propagation (AP)^[1], which will split the target dataset's features into: inlier clusters that each cluster mostly contains inliers (high ranking-index similarity (often larger than 0.9)); outlier clusters that each cluster mostly contains outliers (low ranking-index similarity (often less than 0.5)); mixed clusters that contain both inliers and outliers.

Framework

An unlabeled target dataset is first processed by natural image pretrained feature extractors. We then investigate the intrinsic priors of the feature space (local and global separability), and generate pseudo labels. Based on predicted inliers and outliers, we apply contrastive learning to enhance the discriminativeness of the feature representations.

Experiments

Table 1: Average AUC results on a series of natural image datasets. The best results are highlighted in bold.

Method	Venue	STL-10	Internet	Caltech-101	CIFAR-10	CIFAR-100	MIT-Places
Deep SVDD	ICML-2018	0.593	0.674	0.745	0.533	0.575	0.539
RSRAE	ICLR-2019	0.903	0.916	0.986	0.816	0.889	0.778
GOAD	ICLR-2020	0.946	0.945	0.981	0.862	0.886	0.871
Shell-Re	TPAMI-2021	0.866	0.896	0.905	0.867	0.834	0.822
NeuTraL	ICML-2021	0.828	0.877	0.633	0.748	0.815	0.711
ICL	ICLR-2022	0.929	0.937	0.964	0.859	0.911	0.832
LUNAR	AAAI-2022	0.781	0.776	0.923	0.766	0.859	0.698
ECOD	TKDE-2022	0.932	0.925	0.976	0.888	0.905	0.846
LVAD	ECCV-2022	0.948	0.923	0.977	0.864	0.907	0.860
SLAD	ICML-2023	0.923	0.919	0.888	0.861	0.899	0.846
DeepIF	TKDE-2023	0.858	0.866	0.874	0.801	0.808	0.765
Multi-T	ECCV-2024	0.957	0.956	0.985	0.888	0.917	0.859
FlexUOD	CVPR-2025	0.980	0.981	0.983	0.942	0.947	0.928
VUOD (Ours)	—	0.994	0.991	0.994	0.972	0.981	0.971

Table 2: AUC results on industrial inspection and medical imaging (disease detection).

Type	Dataset	ResNet-18					Wide ResNet-101
Type	Dataset	RSRAE	LVAD	Multi-T	FlexUOD	VUOD	RSRAE	LVAD	Multi-T	FlexUOD	VUOD
Industrial	MVTec-AD	0.709	0.779	0.757	0.752	0.911	0.770	0.799	0.780	0.778	0.931
	BTAD	0.873	0.894	0.885	0.882	0.902	0.846	0.884	0.886	0.883	0.928
	MPDD	0.618	0.573	0.563	0.564	0.906	0.592	0.602	0.603	0.595	0.926
	MVTec-LOCO	0.655	0.693	0.661	0.663	0.812	0.629	0.695	0.657	0.660	0.832
Medical	LiverCT	0.666	0.663	0.696	0.708	0.813	0.703	0.571	0.652	0.567	0.838
	BrainMRI	0.541	0.623	0.583	0.584	0.804	0.676	0.647	0.652	0.578	0.790
	OCT	0.524	0.778	0.730	0.612	0.880	0.713	0.811	0.725	0.588	0.877
	RESC	0.712	0.810	0.823	0.948	0.901	0.662	0.746	0.644	0.589	0.931

3D Dense Point Map Reconstruction Results of VGGT

Plug-and-Play Integrations

We show its plug-and-play property across downstream vision pipelines: simply place VUOD as a data pre-filter before task-specific models.

Image Classification Pipeline

Dataset -> VUOD filtering -> ResNet-18 training

Top-1: 90.81% -> 91.45% | Top-5: 99.11% -> 99.39%

VUOD removes noisy outliers before supervised optimization, improving robustness with minimal integration overhead.

3D Reconstruction Pipeline (VGGT)

Input sequence -> VUOD outlier rejection -> VGGT reconstruction

Per-image filtering latency: 118 ms -> 27 ms (4.4x speedup)

Compared with RobustVGGT^[2]-style rejection, VUOD is backbone-agnostic and delivers faster, stable preprocessing.

Performance and Runtime Efficiency

Clustering Runtime (10,000 ResNet-50 features)

AP: 84.6s
k-Means (k=10): 9.6s
k-Means (k=100): 61.2s
k-Means (k=#AP): 207.3s

MVTec-AD Efficiency (Wide ResNet-101)

Method	Time(S.)	FLOPs(G)	GPU Mem(MB)
EfficientAD [3]	42.1	75.9	102.0
VUOD	17.2	22.8	39.2

BibTeX

@article{liu2026vuod,
  author    = {Zhonghang Liu, Siyuan Chen, Jingwen Yu, Changshuo Wang, Kunyang Li, Jiangbo Lu},
  title     = {VUOD: A Versatile Unsupervised Outlier Detection Framework for Natural, Industrial, Medical Images and Beyond},
  journal   = {European Conference on Computer Vision},
  year      = {2026},
}

References

Frey, Brendan J. and Delbert Dueck. Clustering by passing messages between data points. Science, 2007.
Han, et al. Emergent outlier view rejection in visual geometry grounded transformers. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026.
Batzner, et al. Efficientad: Accurate visual anomaly detection at millisecond-level latencies. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024.