cs.CVSep 6, 2026

Unified Multi-Layer Subspace Modeling for Cross-Domain OOD Detection

Authors: Gerhard Krumpl, Henning Avenhaus, Horst Possegger

Organizations: Institute of Visual Computing, Graz University of Technology, Austria · KESTRELEYE GmbH, Austria

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Out-of-Distribution (OOD) detection remains a fundamental challenge for neural networks, whose predictions can be overconfident on inputs that deviate from the training distribution. Most post-hoc OOD detection methods derive scores from a single representation level (eg., logits or penultimate features) or combine multiple layers via depth selection or OOD-calibrated weighting. However, because OOD shifts are diverse, the most informative representation level can vary strongly across OOD types and domains, making fixed-layer choices and OOD-calibrated aggregation brittle. In this paper, we propose PRISM (Projected Representation with Intermediate-layer Subspace Modeling), a model-agnostic post-hoc OOD detection method that models a unified multi-layer feature representation rather than aggregating independently scored layers. PRISM fuses intermediate and deep features into a single hierarchical embedding, estimates an in-distribution (ID) principal subspace, and then combines two complementary signals: (i) a class-conditional Mahalanobis distance in the projected subspace and (ii) the residual energy orthogonal to the learned manifold. This simple design avoids OOD-tuned layer weighting while capturing both in-subspace semantic deviations and off-subspace anomalies. Across diverse benchmarks spanning natural images, medical imaging, and industrial visual inspection, PRISM achieves consistent state-of-the-art cross-domain OOD detection performance with a single default configuration across all evaluated domains and architectures. We further show that PRISM incurs minimal inference overhead, making it practical for real-world deployment.

Figures & tables

Figure 1: Comprehensive evaluation of post-hoc OOD detection across natural, medical, and industrial benchmarks. (a) Mean AUROC ( ↑\uparrow ) and FPR95 ( ↓\downarrow ) averaged over all datasets, OOD splits, and architectures. (b) Cross-domain correlation between ImageNet-1k [ 14 ] (OpenOOD [ 83 ] ) and OpenMIBOOD [ 19 ] performance (AUROC). Our PRISM ( orange ) consistently exhibits strong performance across domains.
Figure 2: Comparison of OOD score distribution across layer depths. Histograms represent data sampled from PhaKIR (ID) and Cholec80 (Near-OOD) using a ResNet-18. (a) MDS++ applied to the first ResNet-18 block (shallow features). (b) MDS++ applied to the penultimate layer (deep features). (c) PRISM integrates the full architectural hierarchy (shallow to deep features).
Figure 3: Sensitivity of layer aggregation to OOD validation data. For three benchmark tasks: (a) MIDOG, (b) PhaKIR, and (c) ImageNet-1k, we fit linear aggregation weights over ResNet blocks 1–4 using one OOD dataset as validation and evaluate the resulting model across all OOD test sets. The stacked bar plots (top) show the normalized layer weights learned for each calibration dataset, while the bars (bottom) indicate the corresponding mean FPR95 averaged over all OOD datasets.
MethodOpenOODOpenMIBOODICONIC-444Average
C10C100IN-1kMeanMIDOGPhaKIROASIS-3MeanAlmondWheatKernelsFGMean
MSP [ 27 ]087.97\phantom{0}87.97078.55\phantom{0}78.55077.94\phantom{0}77.94081.49\phantom{0}81.49064.23\phantom{0}64.23039.99\phantom{0}39.99069.23\phantom{0}69.23057.82\phantom{0}57.82075.04\phantom{0}75.04039.86\phantom{0}39.86086.74\phantom{0}86.74090.61\phantom{0}90.61073.06\phantom{0}73.06070.79\phantom{0}70.79
MDS [ 40 ]088.66\phantom{0}88.66063.83\phantom{0}63.83089.83\phantom{0}89.83080.77\phantom{0}80.77078.14\phantom{0}78.14073.66\phantom{0}73.66098.73\phantom{0}98.73083.51\phantom{0}83.51097.00\phantom{0}97.00098.64\phantom{0}98.64096.65\phantom{0}96.65088.96\phantom{0}88.96095.31\phantom{0}95.31086.53\phantom{0}86.53
MDSEns* [ 40 ]091.79\phantom{0}91.79078.24\phantom{0}78.24054.41\phantom{0}54.41074.82\phantom{0}74.8296.8798.49099.89‾\phantom{0}\underline{99.89}98.41093.38\phantom{0}93.38098.44\phantom{0}98.44095.82\phantom{0}95.82087.59\phantom{0}87.59093.81\phantom{0}93.81089.01\phantom{0}89.01
GRAM [ 61 ]091.78\phantom{0}91.78085.00‾\phantom{0}\underline{85.00}085.61\phantom{0}85.61087.46‾\phantom{0}\underline{87.46}089.49\phantom{0}89.49036.46\phantom{0}36.46099.42\phantom{0}99.42075.12\phantom{0}75.12097.77‾\phantom{0}\underline{97.77}098.70\phantom{0}98.70097.82\phantom{0}97.8295.8997.55086.71\phantom{0}86.71
EBO [ 45 ]087.49\phantom{0}87.49081.01\phantom{0}81.01041.68\phantom{0}41.68070.06\phantom{0}70.06066.54\phantom{0}66.54032.94\phantom{0}32.94065.83\phantom{0}65.83055.10\phantom{0}55.10064.50\phantom{0}64.50026.11\phantom{0}26.11084.35\phantom{0}84.35094.09\phantom{0}94.09067.26\phantom{0}67.26064.14\phantom{0}64.14
Table 1: AUROC ( ↑\uparrow ) performance across diverse imaging domains. Comparison of post-hoc OOD detection methods on natural (OpenOOD), medical (OpenMIBOOD), and industrial (ICONIC-444) benchmarks. Values represent the mean AUROC ( %\% ) across the benchmark-specific OOD test sets. Standard backbones are employed: ResNet-18 for CIFAR-10/100 (C10, C100), PhaKIR, and ICONIC-444; ResNet-50 for ImageNet-1k (IN-1k) and MIDOG; and R(2+1)D for OASIS-3. The best and second-best results in each column are highlighted in bold and underlined, respectively. * Results of MDSEns should be interpreted with caution, as discussed in Sec. 3.2 and in prior benchmark analyses [ 19 ] .
Figure 5
Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

BenchmarkID DatasetNear-OODFar-OODExtreme-OOD
OpenOOD [ 83 ]CIFAR-10 [ 34 ]CIFAR-100 [ 34 ] , Tiny ImageNet [ 69 ]MNIST [ 39 ] , SVHN [ 53 ] , Textures [ 13 ] , Places365 [ 85 ]MIDOG [ 7 ] , ICONIC-444 [ 36 ]
CIFAR-100 [ 34 ]CIFAR-10 [ 34 ] , Tiny ImageNet [ 69 ]MNIST [ 39 ] , SVHN [ 53 ] , Textures [ 13 ] , Places365 [ 85 ]MIDOG [ 7 ] , ICONIC-444 [ 36 ]
ImageNet-1k [ 14 ]SSB-Hard [ 73 ] , NINCO [ 11 ]iNaturalist [ 72 ] , Textures [ 13 ] , OpenImage-O [ 74 ]MNIST [ 39 ] , FMNIST [ 76 ]
OpenMIBOOD [ 19 ]MIDOG [ 7 ]other MIDOG domain shifts not in IDCCAgT [ 3 , 6 ] , FNAC 2019 [ 60 ]PhaKIR [ 59 ] , ImageNet [ 14 ] , iNaturalist [ 72 ] , Textures [ 13 ]
PhaKIR [ 59 ]Cholec80 [ 71 ] , EndoSeg15 [ 12 ] , EndoSeg18 [ 2 ]Kvasir [ 30 ] , CATARACTS [ 1 ]MIDOG [ 7 ] , ImageNet [ 14 ] , iNaturalist [ 72 ] , Textures [ 13 ]
OASIS-3 [ 38 ]ATLAS [ 42 ] , BraTS [ 8 , 9 , 49 ] , OASIS-CT [ 38 ]MDS-Heart [ 5 , 62 , 68 ] , CHAOS [ 33 , 32 ]MIDOG [ 7 ] , MNIST [ 39 ] , FMNIST [ 76 ] , Textures [ 13 ]
Appendix
Table 4: Detailed benchmark configuration. For each benchmark and in-distribution (ID) dataset, we list the corresponding near-, far-, and extreme-OOD evaluation sets.
BenchmarkID Dataset / TaskArchitectureCheckpoint Source / Training
OpenOOD [ 83 ]CIFAR-10 [ 34 ]ResNet-18 [ 23 ]Official OpenOOD checkpoint
CIFAR-100 [ 34 ]ResNet-18 [ 23 ]Official OpenOOD checkpoint
ImageNet-1k [ 14 ]ResNet-50 [ 23 ]timm: resnet50.tv2_in1k
Swin-T [ 47 ]torchvision: Swin_T_Weights.IMAGENET1K_V1
ViT-B/16 [ 22 ]timm: vit_base_patch16_224.mae_in1k
OpenMIBOOD [ 19 ]MIDOG [ 7 ]ResNet-50 [ 23 ]Official OpenMIBOOD checkpoint
Appendix
Table 5: Overview of architectures and their corresponding official checkpoint sources or training protocol across all benchmarks.
MethodCal. dataHyperparameter search space
IDOODParam 1RangeParam 2Range
MSP [ 27 ]  ————
MDS [ 40 ]✓ ————
MDSEns [ 40 ]✓✓noise[0, 0.0025, 0.0014, 0.005]——
GRAM [ 61 ]✓ ————
EBO [ 45 ]  temperature[0.1, 0.5, 1.0, 1.5, 2.0]——
Appendix
Table 6: Hyperparameter search space and calibration requirements for the considered OOD detection method.
MethodNearFarExtreme 
CIFAR-100Tiny ImageNetMNISTPlaces365SVHNTexturesICONIC-444MIDOGAverage
AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95
↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow
MSP [ 27 ]087.16\phantom{0}87.16060.60\phantom{0}60.60088.85\phantom{0}88.85056.85\phantom{0}56.85092.53\phantom{0}92.53047.03\phantom{0}47.03089.03\phantom{0}89.03056.15\phantom{0}56.15091.31\phantom{0}91.31053.33\phantom{0}53.33089.94\phantom{0}89.94054.96\phantom{0}54.96081.41\phantom{0}81.41061.25\phantom{0}61.25089.02\phantom{0}89.02061.68\phantom{0}61.68087.97\phantom{0}87.97057.69\phantom{0}57.69
MDS [ 40 ]083.62\phantom{0}83.62069.54\phantom{0}69.54084.85\phantom{0}84.85069.34\phantom{0}69.34089.94\phantom{0}89.94063.84\phantom{0}63.84085.01\phantom{0}85.01068.45\phantom{0}68.45091.37\phantom{0}91.37055.65\phantom{0}55.65092.68\phantom{0}92.68040.55\phantom{0}40.55095.33\phantom{0}95.33029.45\phantom{0}29.45088.68\phantom{0}88.68065.56\phantom{0}65.56088.66\phantom{0}88.66058.02\phantom{0}58.02
MDSEns [ 40 ]084.75\phantom{0}84.75067.18\phantom{0}67.18086.19\phantom{0}86.19066.71\phantom{0}66.71093.94\phantom{0}93.94043.02\phantom{0}43.02086.15\phantom{0}86.15067.42\phantom{0}67.42091.30\phantom{0}91.30052.09\phantom{0}52.09091.69\phantom{0}91.69040.01\phantom{0}40.01100.000.00098.28\phantom{0}98.28010.08\phantom{0}10.08091.79\phantom{0}91.79040.87\phantom{0}40.87
Appendix
Table 7: Performance of OOD detection methods for a ResNet-18 backbone trained on CIFAR-10.
MethodNearFarExtreme 
CIFAR-10Tiny ImageNetMNISTPlaces365SVHNTexturesICONIC-444MIDOGAverage
AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95
↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow
MSP [ 27 ]078.49\phantom{0}78.49079.85\phantom{0}79.85082.12\phantom{0}82.12074.53\phantom{0}74.53075.64\phantom{0}75.64087.65\phantom{0}87.65079.27\phantom{0}79.27079.31\phantom{0}79.31079.02\phantom{0}79.02078.67\phantom{0}78.67077.57\phantom{0}77.57080.73\phantom{0}80.73083.35\phantom{0}83.35074.86\phantom{0}74.86071.59\phantom{0}71.59090.70\phantom{0}90.70078.55\phantom{0}78.55080.52\phantom{0}80.52
MDS [ 40 ]056.00\phantom{0}56.00095.91\phantom{0}95.91061.57\phantom{0}61.57094.29\phantom{0}94.29067.97\phantom{0}67.97093.46\phantom{0}93.46063.18\phantom{0}63.18093.24\phantom{0}93.24070.60\phantom{0}70.60090.24\phantom{0}90.24076.13\phantom{0}76.13073.09\phantom{0}73.09068.61\phantom{0}68.61087.86\phantom{0}87.86057.87\phantom{0}57.87097.63\phantom{0}97.63063.83\phantom{0}63.83091.78\phantom{0}91.78
MDSEns [ 40 ]057.16\phantom{0}57.16095.38\phantom{0}95.38065.66\phantom{0}65.66092.47\phantom{0}92.47088.55\phantom{0}88.55061.05\phantom{0}61.05067.00\phantom{0}67.00091.07\phantom{0}91.07083.36\phantom{0}83.36071.00\phantom{0}71.00085.88\phantom{0}85.88051.37\phantom{0}51.37093.00\phantom{0}93.00007.36\phantom{0}\phantom{0}7.36091.24‾\phantom{0}\underline{91.24}047.92‾\phantom{0}\underline{47.92}078.24\phantom{0}78.24063.40\phantom{0}63.40
Appendix
Table 8: Performance of OOD detection methods for a ResNet-18 backbone trained on CIFAR-100.
 MethodNearFarExtreme 
 SSB-HardNINCOiNaturalistTexturesOpenImage-OMNISTFashionMNISTAverage
 AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95
 ↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow
ResNet-50MSP [ 27 ]068.32\phantom{0}68.32085.50\phantom{0}85.50076.73\phantom{0}76.73076.15\phantom{0}76.15084.80\phantom{0}84.80059.68\phantom{0}59.68077.34\phantom{0}77.34070.50\phantom{0}70.50082.29\phantom{0}82.29067.07\phantom{0}67.07078.39\phantom{0}78.39078.01\phantom{0}78.01081.22\phantom{0}81.22062.81\phantom{0}62.81077.94\phantom{0}77.94072.33\phantom{0}72.33
MDS [ 40 ]070.44\phantom{0}70.44081.32\phantom{0}81.32082.58\phantom{0}82.58070.13\phantom{0}70.1395.2526.25093.57\phantom{0}93.57026.16\phantom{0}26.1692.14043.12‾\phantom{0}\underline{43.12}099.75\phantom{0}99.75000.03\phantom{0}\phantom{0}0.03098.90\phantom{0}98.90003.13\phantom{0}\phantom{0}3.13089.83\phantom{0}89.83036.38‾\phantom{0}\underline{36.38}
MDSEns [ 40 ]043.20\phantom{0}43.20097.98\phantom{0}97.98051.52\phantom{0}51.52095.82\phantom{0}95.82058.17\phantom{0}58.17097.52\phantom{0}97.52075.37\phantom{0}75.37065.95\phantom{0}65.95056.26\phantom{0}56.26091.98\phantom{0}91.98054.65\phantom{0}54.65100.00100.00050.58\phantom{0}50.58099.94\phantom{0}99.94054.41\phantom{0}54.41094.01\phantom{0}94.01
Appendix
Table 9: Performance of OOD detection methods on ImageNet-1k using three backbones: ResNet-50, ViT-B/16, and Swin-T. Metrics are reported in %\% (higher is better for AUROC; lower is better for FPR95, indicated by ↑/↓\uparrow/\downarrow ). The best and second-best results for each OOD dataset ( i.e . , each column) are shown in bold or underlined, respectively.
 MethodNearFarExtreme 
 MeanMeanImageNet-1kiNaturalistTexturesPlaces365MeanAverage
 AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95
 ↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow
ResNet-18MSP [ 27 ]072.86\phantom{0}72.86072.12\phantom{0}72.12074.85\phantom{0}74.85064.22\phantom{0}64.22076.06\phantom{0}76.06055.12\phantom{0}55.12064.79\phantom{0}64.79062.83\phantom{0}62.83068.46\phantom{0}68.46064.08\phantom{0}64.08076.59\phantom{0}76.59053.52\phantom{0}53.52071.48\phantom{0}71.48058.89\phantom{0}58.89073.06\phantom{0}73.06065.07\phantom{0}65.07
MDS [ 40 ]089.14\phantom{0}89.14032.15\phantom{0}32.15097.03\phantom{0}97.03014.25\phantom{0}14.25099.73\phantom{0}99.73000.54\phantom{0}\phantom{0}0.54099.86\phantom{0}99.86000.15\phantom{0}\phantom{0}0.15099.72\phantom{0}99.72000.99\phantom{0}\phantom{0}0.99099.74\phantom{0}99.74000.45\phantom{0}\phantom{0}0.45099.76\phantom{0}99.76000.53\phantom{0}\phantom{0}0.53095.31\phantom{0}95.31015.65\phantom{0}15.65
MDSEns [ 40 ]087.94\phantom{0}87.94028.82\phantom{0}28.82097.15\phantom{0}97.15010.01\phantom{0}10.01096.65\phantom{0}96.65003.63\phantom{0}\phantom{0}3.63098.34\phantom{0}98.34001.76\phantom{0}\phantom{0}1.76093.34\phantom{0}93.34006.99\phantom{0}\phantom{0}6.99097.04\phantom{0}97.04003.29\phantom{0}\phantom{0}3.29096.34\phantom{0}96.34003.92\phantom{0}\phantom{0}3.92093.81\phantom{0}93.81014.25\phantom{0}14.25
Appendix
Table 10: Average OOD detection performance per method and OOD test set using three backbones: Resnet-18, CCT-7/7x2, and ConvNeXt-S. Results are reported as the mean over the four tasks (Almond, Wheat, Kernels, and Food-grade). Arrows ( ↑\uparrow / ↓\downarrow ) indicate whether higher or lower values are better. Additionally, the best and second-best results in each column are highlighted in bold and underlined, respectively. All values are reported as percentages.
Figure 6: Mean-rank comparison (AUROC) across OOD datasets. Average rank of OOD detection methods computed over OOD datasets. Error bars show 95%95\% CIs; lower rank is better.
Figure 7: Qualitative analysis for OpenOOD (ImageNet-1k / SSB-hard). For each column, the top row shows an OOD sample from the SSB-hard split, with its original OOD class label annotated in blue (parentheses). The two rows below show two ID ImageNet examples from the ID class predicted by the Mahalanobis component of our detector ( i.e . , the class achieving the minimum Mahalanobis distance in PCA space), with the corresponding predicted ID class label annotated in red .
Figure 8: Qualitative analysis for OpenOOD (ImageNet-1k / Textures). For each column, the top row shows an OOD sample from the Textures split, with its original OOD class label annotated in blue (parentheses). The two rows below show two ID ImageNet examples from the ID class predicted by the Mahalanobis component of our detector ( i.e . , the class achieving the minimum Mahalanobis distance in PCA space), with the corresponding predicted ID class label annotated in red .
Figure 9: Qualitative analysis for OpenOOD (ImageNet-1k / NINCO). For each column, the top row shows an OOD sample from the NINCO split, with its original OOD class label annotated in blue (parentheses). The two rows below show two ID ImageNet examples from the ID class predicted by the Mahalanobis component of our detector ( i.e . , the class achieving the minimum Mahalanobis distance in PCA space), with the corresponding predicted ID class label annotated in red .
Figure 10: Qualitative analysis on OpenMIBOOD/PhaKIR. For each column, the top row shows an OOD sample annotated with the source OOD dataset ( blue ). The two rows below show two ID examples from the Mahalanobis-predicted ID class ( red ; argmin Mahalanobis distance in PCA space), illustrating which ID category each OOD sample is mapped to.
Figure 11: ICONIC-444/almond qualitative examples. For each column, the top row shows an OOD sample with its OOD class label ( blue ), and the two rows below show two ID examples from the Mahalanobis-predicted ID class ( red ; argmin Mahalanobis distance in PCA space).
MethodMIDOGPhaKIROASIS-3 
NearFarExtremeMeanNearFarExtremeMeanNearFarExtremeMeanAverage
AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95
↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow↑\uparrow↓\downarrow
MSP [ 27 ]055.90\phantom{0}55.90090.51\phantom{0}90.51080.91\phantom{0}80.91071.45\phantom{0}71.45055.87\phantom{0}55.87091.13\phantom{0}91.13064.23\phantom{0}64.23084.36\phantom{0}84.36050.14\phantom{0}50.14096.21\phantom{0}96.21032.50\phantom{0}32.50097.65\phantom{0}97.65037.32\phantom{0}37.32097.96\phantom{0}97.96039.99\phantom{0}39.99097.27\phantom{0}97.27054.56\phantom{0}54.56093.62\phantom{0}93.62079.16\phantom{0}79.16089.17\phantom{0}89.17073.97\phantom{0}73.97083.24\phantom{0}83.24069.23\phantom{0}69.23088.68\phantom{0}88.68057.82\phantom{0}57.82090.10\phantom{0}90.10
MDS [ 40 ]063.21\phantom{0}63.21081.00\phantom{0}81.00090.91\phantom{0}90.91041.60\phantom{0}41.60080.30\phantom{0}80.30062.95\phantom{0}62.95078.14\phantom{0}78.14061.85\phantom{0}61.85076.49\phantom{0}76.49070.30\phantom{0}70.30051.48\phantom{0}51.48091.14\phantom{0}91.14093.02\phantom{0}93.02021.38\phantom{0}21.38073.66\phantom{0}73.66060.94\phantom{0}60.94096.19\phantom{0}96.19030.57\phantom{0}30.57100.000.00100.000.00098.73\phantom{0}98.73010.19\phantom{0}10.19083.51\phantom{0}83.51044.33\phantom{0}44.33
MDSEns* [ 40 ]90.6021.83100.000.00100.000.0096.877.2897.0521.2498.4115.85100.000.0098.4912.36099.67‾\phantom{0}\underline{99.67}002.31\phantom{0}\phantom{0}2.31100.000.00100.000.00099.89‾\phantom{0}\underline{99.89}000.77\phantom{0}\phantom{0}0.7798.416.80
Appendix
Table 11: Average OOD detection performance for the OpenMIBOOD benchmark, measured as FPR at 95%95\% TPR, reported per method and OOD category (near, far, and extreme. Arrows ( ↑\uparrow / ↓\downarrow ) indicate whether higher or lower values are better. Additionally, the best and second-best results in each column are highlighted in bold and underlined, respectively. All values are reported as percentages. *: Results of MDSEns should be interpreted with caution, as discussed in the main manuscript and in prior benchmark analyses [ 19 ] .
More in this topic

OOD Detection

Out-of-distribution (OOD) detection identifies inputs unlike a model’s training data so systems can flag uncertainty, reject inputs, or adapt safely.

See all 215 papers

Explore similar work

CardsList
Mar 24, 2026

Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection

Deep learning models are increasingly deployed in safety-critical applications, where reliable out-of-distribution (OOD) detection is essential to ensure robustness. Existing methods predominantly rely on the penultimate-layer activations of neural networks, assuming they encapsulate the most informative in-distribution (ID) representations. In this work, we revisit this assumption to show that intermediate layers encode equally rich and discriminative information for OOD detection. Based on this observation, we propose a simple yet effective model-agnostic approach that leverages internal representations across multiple layers. Our scheme aggregates features from successive convolutional blocks, computes class-wise mean embeddings, and applies L_2 normalization to form compact ID prototypes capturing class semantics. During inference, cosine similarity between test features and these prototypes serves as an OOD score--ID samples exhibit strong affinity to at least one prototype, whereas OOD samples remain uniformly distant. Extensive experiments on state-of-the-art OOD benchmarks across diverse architectures demonstrate that our approach delivers robust, architecture-agnostic performance and strong generalization for image classification. Notably, it improves AUROC by up to 4.41% and reduces FPR by 13.58%, highlighting multi-layer feature aggregation as a powerful yet underexplored signal for OOD detection, challenging the dominance of penultimate-layer-based methods. Our code is available at: https://github.com/sgchr273/cosine-layers.git.

Shreen Gul, Mohamed Elmahallawy, Ardhendu Tripathy +1

Missouri University of Science and Technology, Rolla, MO 65401, USA · Washington State University, Richland, WA 99354, USA

OOD Detection
AskPDF
Jun 15, 2026

MM++: Post-Hoc Scale-Invariant Multilayer OOD Detection via Top-K Gated Feature Fusion

We introduce MM++ (Multilayer Mahalanobis++), a strictly post-hoc, and scale-invariant framework for Out-of-Distribution (OOD) detection. To address the trade-off between scale invariance and hierarchical expressivity, MM++ constructs a principled joint feature space. It first identifies discriminative intermediate layers by measuring entropy density drops, which mark the boundaries of sharp semantic compression. By fusing these selected layers with the terminal representation, the framework captures latent cross-layer correlations while mitigating early-layer noise. Crucially, a Ledoit-Wolf regularized tied covariance matrix stabilizes this unified space, enabling reliable distance estimation. Requiring no auxiliary OOD data, classifier fine-tuning, or architectural modifications, MM++ delivers robust performance across distinct architectures for both near- and far-OOD detection.

Rahim Hossain, Md Tawheedul Islam Bhuian, Md Farhan Shadiq +1

School of Computing State University of New York at Binghamton

Diffusion-Based OOD DetectionOOD Detection
AskPDF
May 5, 2026

Geometry over Density: Few-Shot Cross-Domain OOD Detection

Out-of-distribution (OOD) detection identifies test samples that fall outside a model's training distribution, a capability critical for safe deployment in high-stakes applications. Standard OOD detectors are trained on a specific in-distribution (ID) dataset and detect deviations from that single domain. In contrast, we study few-shot cross-domain OOD detection: given a single pre-trained model, can we perform OOD detection on arbitrary new ID-OOD task pairs using only a handful of ID samples at inference time, with no additional training? We propose UFCOD, a unified framework that achieves this goal through information-geometric analysis of diffusion trajectories. Our key insight is that diffusion noise predictions are score functions (gradients of log-density), and we extract two energy features: Path Energy (integrated score magnitude) and Dynamics Energy (score smoothness), that form a discrete Sobolev norm capturing how samples interact with the learned diffusion process. The central contribution is a train-once, deploy-anywhere paradigm: a diffusion model trained on a single dataset (e.g., CelebA) serves as a universal feature extractor for OOD detection across semantically unrelated domains (e.g., CIFAR-10, SVHN, Textures). At deployment, each new task requires only ∼\sim100 unlabeled ID samples for inference: no retraining, no fine-tuning, no task-specific adaptation. Using 100 ID samples per task, UFCOD achieves 93.7% average AUROC across 12 cross-domain benchmarks, competitive with methods trained on 50k--163k samples, demonstrating ∼\sim500×\times improvement in sample efficiency. See our code in https://github.com/lili0415/UFCOD.

Shawn Li, You Qin, Jiate Li +4

University of Southern California · National University of Singapore · Amazon

Diffusion-Based OOD DetectionDistribution Shift DetectionCross-Domain GeneralizationOOD Detection
AskPDF