Detectors¶

Out-of-Distribution detectors, grouped by method family. All of them implement the common detector interface: fit on in-distribution data (only where a detector requires it), then call the detector to get outlier scores, which are larger for outliers.

Comparison¶

Detector	Input	Fit	Classification	Segmentation	Paper
`ASH`	feature maps		✓		ICLR 2023
`DICE`	features	yes	✓		ECCV 2022
`EnergyBased`	logits		✓	✓	NeurIPS 2020
`Entropy`	logits		✓	✓	ICCV 2021
`fDBD`	features	yes	✓		ICML 2024
`GEN`	logits		✓	✓	CVPR 2023
`GMM`	features	yes	✓		—
`GradNorm`	model inputs		✓		NeurIPS 2021
`GradUncertainty`	model inputs		✓		ICIP 2020
`Gram`	several layers	yes	✓		ICML 2020
`KLMatching`	logits	yes	✓		ICML 2022
`KNN`	features	yes	✓		ICML 2022
`LTS`	features		✓		Machine Vision and Applications 2025
`Mahalanobis`	features	yes	✓		NeurIPS 2018
`MahalanobisODIN`	model inputs	yes	✓		NeurIPS 2018
`MaxLogit`	logits		✓	✓	ICML 2022
`MaxSoftmax`	logits		✓	✓	ICLR 2017
`MCD`	model inputs		✓	✓	ICML 2016
`MCM`	features		✓		NeurIPS 2022
`MultiMahalanobis`	several layers	yes	✓		NeurIPS 2018
`NACUE`	model inputs	yes	✓		ICLR 2024
`NCI`	features	yes	✓		CVPR 2025
`NNGuide`	features	yes	✓		ICCV 2023
`ODIN`	model inputs		✓		ICLR 2018
`OpenMax`	logits	yes	✓		CVPR 2016
`PNML`	features	yes	✓		NeurIPS 2021
`RankFeat`	feature maps		✓		NeurIPS 2022
`ReAct`	feature maps	yes	✓	✓	NeurIPS 2021
`RMD`	features	yes	✓		arXiv 2021
`SCALE`	feature maps		✓		ICLR 2024
`SHE`	features	yes	✓		ICLR 2023
`TemperatureScaling`	logits	yes	✓		ICML 2017
`ViM`	features	yes	✓		CVPR 2022
`VRA`	feature maps	yes	✓		NeurIPS 2023
`WeightedEBO`	logits		✓	✓	ICLR 2022

Probability-based¶

Probability-based methods are based on the observation that OOD inputs tend to be assigned lower posteriors with higher entropy, i.e., the predicted distribution is often less concentrated on a single class.

Logit-based¶

Logit-based methods are based on the observation that OOD inputs tend to yield different logits compared to ID data.

Feature-based¶

Gradient-based¶

Gradient-based detectors are based on the observation that the gradients (w.r.t. the model parameters, the inputs or intermediate activations) for ID and OOD data behave differently. All gradient-based detectors inherit from pytorch_ood.api.GradientDetector.

Activation Pruning¶

Activation pruning methods are based on the observation that OOD inputs cause unusual activations in the model, and that, by rectifying these unusual activations, we can often improve discriminability of ID and OOD samples.