Mitigating the ID-OOD Tradeoff in Open-Set Test-Time Adaptation

Wenjie Zhao; Jia Li; Xin Dong; Yapeng Tian; Yu Xiang; Yunhui Guo

cs.CVApr 2, 2026

Mitigating the ID-OOD Tradeoff in Open-Set Test-Time Adaptation

Authors: Wenjie Zhao, Jia Li, Xin Dong, Yapeng Tian, Yu Xiang, Yunhui Guo

Organizations: University of Texas at Dallas · NVIDIA

Abstract

Open-set test-time adaptation (OSTTA) addresses the challenge of adapting models to new environments where out-of-distribution (OOD) samples coexist with in-distribution (ID) samples affected by distribution shifts. In such settings, covariate shift-for example, changes in weather conditions such as snow-can alter ID samples, reducing model reliability. Consequently, models must not only correctly classify covariate-shifted ID (csID) samples but also effectively reject covariate-shifted OOD (csOOD) samples. Entropy minimization is a common strategy in test-time adaptation to maintain ID performance under distribution shifts, while entropy maximization is widely applied to enhance OOD detection. Several studies have sought to combine these objectives to tackle the challenges of OSTTA. However, the intrinsic conflict between entropy minimization and maximization inevitably leads to a trade-off between csID classification and csOOD detection. In this paper, we first analyze the limitations of entropy maximization in OSTTA and then introduce an angular loss to regulate feature norm magnitudes, along with a feature-norm loss to suppress csOOD logits, thereby improving OOD detection. These objectives form ROSETTA, a $\underline{r}$ obust $\underline{o}$ pen- $\underline{se}$ t $\underline{t}$ est- $\underline{t}$ ime $\underline{a}$ daptation. Our method achieves strong OOD detection while maintaining high ID classification performance on CIFAR-10-C, CIFAR-100-C, Tiny-ImageNet-C and ImageNet-C. Furthermore, experiments on the Cityscapes validate the method's effectiveness in real-world semantic segmentation, and results on the HAC dataset demonstrate its applicability across different open-set TTA setups.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Examples of csID and csOOD samples. Open-set TTA adapts the model to handle both simultaneously. (b) Comparison with recent state-of-the-art methods. Our method achieves strong performance on both tasks.

Method	CIFAR-10-C	CIFAR-100-C	TinyImageNet-C	ImageNet-C
GMM classifier	74.73	70.74	59.23	55.42
Entropy K-means	70.16	67.73	56.79	58.33
Energy K-means	73.92	65.86	56.78	55.00
Entropy (Best Thre.)	75.98	70.15	59.90	67.08
Energy (Best Thre.)	75.43	68.79	60.54	65.42

Table 1: OOD detector accuracy comparison on CIFAR-10, CIFAR-100, TinyImageNet and ImageNet datasets. All methods show a non-negligible amount of misclassification.

Figure 2: Examples of misclassification under the optimal entropy threshold. Green boxes represent correctly detected samples and red boxes represent misclassified ones, where csID images are often mistaken for csOOD.

Panel of Figure 3 — Figure 3: $l_{2}$ -norm of feature vectors on CIFAR-10-C under various corruption types. The x-axis corresponds to corruption types, and the y-axis indicates the $l_{2}$ -norm of features

Figure 4: Overview of the ROSETTA pipeline.

Panel of Figure 5 — Figure 5: Mean sorted logits under different adaptation methods.

Method	CIFAR-10-C				CIFAR-100-C				Tiny-ImageNet-C				ImageNet-C
Method	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$
Source	81.73	77.89	79.45	68.44	53.25	60.55	94.98	39.87	22.29	53.79	93.41	16.29	28.21	49.63	94.74	19.81
BN Adapt	84.20	80.40	76.84	72.13	57.16	72.45	84.29	47.10	37.00	61.06	90.90	28.50	43.57	55.89	93.39	30.42
CoTTA	85.77	85.89	72.40	77.26	56.46	77.04	80.96	48.95	12.68	41.36	94.43	7.61	47.67	55.58	94.51	33.8
TENT	79.38	65.39	95.94	56.73	54.74	65.00	94.79	42.24	28.96	49.78	95.96	19.02	45.82	51.34	96.47	30.33
+ UniEnt	84.31	92.28	36.74	80.32	59.07	89.28	51.14	56.26	37.23	63.92	89.72	30.18	47.53	56.33	95.21	34.42

Table 2: Results of the baseline methods and the proposed ROSETTA (Ours) on CIFAR-C benchmarks and ImageNet-C benchmarks. FPR95 indicates the False Positive Rate when the True Positive Rate is 95%. The results demonstrate that our method achieves strong performance in both csID classification and csOOD detection simultaneously.

Based Method	$\mathcal{L}_{\text{csID}}$	$\mathcal{L}_{\text{ang}}$	$\mathcal{L}_{\text{norm}}$	CIFAR-10-C				CIFAR-100-C
Based Method	$\mathcal{L}_{\text{csID}}$	$\mathcal{L}_{\text{ang}}$	$\mathcal{L}_{\text{norm}}$	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$
TENT+Ours	✗	✗	✗	79.38	65.39	95.94	56.73	54.74	65.00	94.79	42.24
	✓	✗	✗	85.04	81.80	68.89	73.57	59.30	86.09	63.65	55.55
	✓	✓	✗	85.66	85.63	65.67	77.32	60.88	85.97	65.05	56.90
	✓	✓	✓	84.34	93.75	29.91	81.37	59.20	91.80	35.89	56.80
EATA+Ours	✗	✗	✗	80.92	84.32	71.66	72.63	60.63	88.64	50.18	57.24

Table 3: Comparison of methods on CIFAR-10-C and CIFAR-100-C datasets w/ and w/o

\mathcal{L}_{\text{csID}}

\mathcal{L}_{\text{ang}}

, and the

l_{1}

-norm loss

\mathcal{L}_{\text{norm}}

Condition	Fog	Night	Rain	Snow
TENT	69.17	40.70	58.68	57.16
+UniEnt	69.13 (-0.04)	40.60 (-0.10)	58.68 (+0.00)	57.02 (-0.14)
+UniEnt+	69.15 (-0.02)	40.64 (-0.06)	58.59 (-0.09)	56.99 (-0.17)
+Ours	69.24 (+0.07)	40.71 (+0.01)	58.92 (+0.24)	57.41 (+0.25)

Table 4: Semantic segmentation results (mIoU) on the Cityscapes-to-ACDC test-time adaptation task.

	H $\rightarrow$ A				H $\rightarrow$ C				C $\rightarrow$ A				C $\rightarrow$ H
Method	Acc↑	FPR95↓	AUROC↑	H-score↑	Acc↑	FPR95↓	AUROC↑	H-score↑	Acc↑	FPR95↓	AUROC↑	H-score↑	Acc↑	FPR95↓	AUROC↑	H-score↑
Source	57.28	87.53	45.90	25.12	38.79	98.99	22.49	2.83	67.88	65.23	70.81	52.07	61.36	71.67	67.72	45.21
Tent	57.51	78.59	65.63	37.82	41.64	94.03	47.25	14.11	66.78	80.68	63.56	36.38	62.94	78.73	65.22	38.35
SAR	58.50	77.70	68.39	39.19	46.32	95.04	48.67	12.31	66.67	63.91	70.11	52.66	66.91	74.41	66.88	43.49
OSTTA	55.63	76.82	72.14	40.01	44.49	89.98	58.36	21.52	65.56	73.73	64.03	43.52	63.52	78.30	64.44	38.79
UniEnt	60.15	78.04	64.02	38.57	40.99	98.71	31.86	3.61	67.55	54.53	79.28	60.72	66.26	64.38	76.11	53.28

Table 5: Multimodal open-set TTA on the HAC dataset with video and audio modalities. Our method achieves strong performance across diverse domains.

Panel of Figure 6 — Figure 6: OSCR heatmaps of ROSETTA with TENT adaptation across hyperparameter settings on CIFAR-10-C and CIFAR-100-C, demonstrating consistent robustness under varying configurations.

Figure 7: Effect of increasing OOD loss weight $\tau$ on ID classification and OOD detection. Our method achieves balanced and superior performance on both tasks.

Method	0.2	0.4	0.6	0.8	1.0	$\Delta\downarrow$	Mean
Source	40.00	40.03	39.98	39.92	39.87	0.16	39.96
BN Adapt	49.91	49.55	48.92	47.97	47.10	2.81	48.69
TENT	47.68	44.12	44.06	42.90	42.16	5.52	44.18
+ UniEnt	56.84	57.48	57.13	56.77	56.26	1.22	56.90
+ UniEnt+	57.15	57.59	57.24	56.88	56.33	1.26	57.04
+ Ours	56.63	57.51	57.51	57.13	56.80	0.88	57.12

Table 6: OSCR performance under different csOOD-to-csID imbalance ratios on CIFAR-100-C. The

\Delta

column indicates the performance variation, calculated as the difference between the maximum and minimum values across the ratios. ROSETTA achieves the best overall accuracy and the most consistent performance across all imbalance settings.

Method	Cifar-10-C	Cifar-100-C	Tiny-ImageNet-C	ImageNet-C
TENT	85.85	62.86	33.13	45.93
+UniEnt	83.26 (-2.59)	60.01 (-2.85)	36.60 (+3.47)	44.48 (-1.45)
+UniEnt+	83.93 (-1.92)	59.73 (-3.13)	35.98 (+2.85)	39.96 (-5.97)
+Ours	84.01 (-1.84)	60.39 (-2.47)	36.62 (+3.49)	46.99 (+1.06)

Table 7: Experiments without csOOD data. Our method is more robust than UniEnt and UniEnt+, which alternate in yielding their highest accuracy. Tent occasionally achieves higher Acc since it does not account for csOOD.

Method	2	4	6	8	10	$\Delta\downarrow$	Mean
Source	70.84	69.28	69.32	69.18	68.44	2.40	69.41
BN Adapt	72.56	72.48	72.52	72.44	72.14	0.42	72.43
TENT	49.51	48.29	51.74	49.53	50.97	3.45	50.01
+ UniEnt	78.71	78.39	78.28	78.13	77.82	0.89	78.27
+ UniEnt+	78.65	78.23	78.23	78.07	77.68	0.97	78.17
+ Ours	81.50	81.37	81.84	81.87	81.37	0.50	81.59

Table 8: OSCR performance with increasing numbers of unknown classes in the csOOD set. As the number of unknown classes grows from 2 to 10, the task becomes more challenging. ROSETTA maintains consistently strong results and shows minimal performance degradation, as reflected by the lowest

\Delta

value.

Dataset	OOD Score	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$
CIFAR-10-C	Energy	84.34	93.75	29.91	81.37
	$l_{1}$ -norm	84.34	94.01	25.19	81.04
	NAN	84.34	92.87	35.88	80.50
CIFAR-100-C	Energy	59.20	91.80	35.89	56.80
	$l_{1}$ -norm	59.20	92.40	29.89	56.01
	NAN	59.20	91.95	31.91	55.88

Table 9: OOD detection performance comparison across different OOD scores on CIFAR-10-C, CIFAR-100-C, and Tiny-ImageNet-C datasets.

Corruption	Acc $\uparrow$	AUROC $\uparrow$	FPR95 $\downarrow$	OSCR $\uparrow$
Gaussian noise	37.99	48.25	95.41	18.54
Shot noise	38.83	52.32	93.90	24.32
Impulse noise	30.75	51.41	94.62	20.19
Defocus blur	38.89	58.68	92.09	27.52
Glass blur	28.96	58.14	91.67	20.76
Motion blur	43.60	65.47	88.92	34.01

Table 10: Performance metrics for each task using

l_{1}

-norm as the OOD score of our method on Tiny-ImageNet-C, ordered by task sequence during adaptation. An AUROC below 50 indicates a tendency to classify OOD samples as ID more frequently than random guessing.

Panel of Figure 8 — Figure 8: T-SNE Visualization on CIFAR-10-C Test Set with SVHN-C as csOOD. Black markers represent csOOD samples, while colored markers represent csID samples. Circles indicate correctly classified csID samples or correctly detected csOOD samples, whereas triangles denote misclassified csID samples or undetected csOOD samples. Stars represent the embeddings of the updated class prototypes, and diamonds represent the fixed weights of the linear classification layer.

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 322 papers

Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation
Apr 23, 2026Yingkai Yang, Chaoqi Chen, Hui HuangDomain Adaptation Continual Test-Time Adaptation
College of Computer Science and Software Engineering, Shenzhen University
A Closer Look at In-Distribution vs. Out-of-Distribution Accuracy for Open-Set Test-time Adaptation
Jun 1, 2026Zefeng Li, Evan ShelhamerTest-Time Adaptation Open-Set Recognition
University of British Columbia and Vector Institute
Dual Strategies for Test-Time Adaptation
Apr 19, 2026Nam Nguyen Phuong, Duc Nguyen The Minh, Phi Le Nguyen +2Distribution Shift Robustness Test-Time Adaptation
Institute for AI Innovation and Societal Impact, Hanoi University of Science and Technology · Monash University · Adelaide University

Apr 23, 2026J/K move · Enter open · S save

Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation

Yingkai Yang, Chaoqi Chen, Hui Huang

College of Computer Science and Software Engineering, Shenzhen University

Domain Adaptation Continual Test-Time Adaptation OOD Detection

Open paper AskPDF arXiv

Test-Time Adaptation (TTA) aims to mitigate distributional shifts between training and test domains during inference time. However, existing TTA methods fall short in the realistic scenario where models face both continually changing domains and the simultaneous emergence of unknown semantic classes, a challenging setting we term Open-set Continual Test-Time Adaptation (OCTTA). The coupling of domain and semantic shifts often collapses the feature space, severely degrading both classification and out-of-distribution detection. To tackle this, we propose DOmain COmpensation (DOCO), a lightweight and effective framework that robustly performs domain adaptation and OOD detection in a synergistic, closed loop. DOCO first performs dynamic, adaptation-conditioned sample splitting to separate likely ID from OOD samples. Then, using only the ID samples, it learns a domain compensation prompt by aligning feature statistics with the source domain, guided by a structural preservation regularizer that prevents semantic distortion. This learned prompt is then propagated to the OOD samples within the same batch, effectively isolating their semantic novelty for more reliable detection. Extensive experiments on multiple challenging benchmarks demonstrate that DOCO outperforms prior CTTA and OSTTA methods, establishing a new state-of-the-art for the demanding OCTTA setting.

Abstract

Test-Time Adaptation

Explore similar work

Back to Source: Open-Set Continual Test-Time Adaptation via Domain Compensation

A Closer Look at In-Distribution vs. Out-of-Distribution Accuracy for Open-Set Test-time Adaptation

Dual Strategies for Test-Time Adaptation