acceptodds
Under review as a conference paper at ICLR 2027

PROBE: Protected Observation-Subspace Energy for Fine-Grained Open-Set Recognition on Frozen Classifiers

Abstract

Fine-grained open-set recognition must reject unseen categories that lie close to known classes while tolerating observation-induced variation in the features. In existing post-hoc scores, feature magnitude varies with prediction confidence and image conditions, and within-class covariance alone does not distinguish observation-induced variation from directions that separate known-class means. We introduce Protected Observation-Subspace Energy (PROBE), which combines class-specific isotropic energies on unit-normalized features with an additive covariance adjustment constructed from paired feature displacements under mild image corruptions, in which a between-class penalty discourages additional tolerance along directions separating known-class means. Among post-hoc and end-to-end baselines, PROBE achieves the highest mean Hard AUROC and Hard open-set classification rate (OSCR) on three Semantic Shift Benchmark (SSB) datasets and the highest mean AUROC and OSCR on three crop-level unmanned aerial vehicle (UAV) protocols constructed here. With the covariance spectrum held fixed, directions fitted with the between-class penalty yield higher mean AUROC than directions fitted without it on every dataset. Fitting uses only known-class samples of a frozen classifier and has a closed-form solution, and inference scores a single test image without corrupting it, retraining the classifier, or changing its label predictions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fine-grained open-set recognition, Protected observation subspace, Class-conditional energy, Post-hoc rejection, Gaussian energy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_stRpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PROBE, a post-hoc open-set recognition method for frozen classifiers, with a particular focus on fine-grained recognition. The method combines two components: (1) class-conditional isotropic Gaussian energies computed from unit-normalized features, and (2) an observation-aware covariance adjustment estimated from paired feature displacements induced by mild image corruptions. To avoid excessively enlarging class support along discriminative directions, the authors introduce a between-class scatter penalty when estimating the observation subspace. The resulting rejection score is computed using a class-conditional energy function without retraining the classifier or modifying its predictions.

Critical strengths
  1. The paper addresses an important challenge in fine-grained open-set recognition: distinguishing observation-induced feature variations from semantic differences between visually similar categories. The motivation is intuitive, and the discussion of why uniformly expanding known-class support may negatively affect unknown-class rejection is reasonable.
  2. PROBE does not require retraining the backbone or classification head, and the subspace estimation has a closed-form solution.
Critical weaknesses
  1. The proposed method primarily combines existing techniques, including feature normalization, Gaussian energy modeling, perturbation-based subspace estimation, and between-class scatter regularization. In particular, the penalized eigendecomposition of (S_{\mathrm{obs}}-\beta S_B) and the subsequent covariance adjustment appear to be relatively straightforward extensions of established ideas. The paper does not sufficiently clarify its fundamental methodological novelty beyond this combination.
  2. PROBE assumes that directions separating known-class means can serve as a proxy for directions important for rejecting unknown classes. However, this assumption is not necessarily valid, especially in fine-grained settings where unseen categories may differ along directions not captured by known-class scatter. While the ablations demonstrate the effectiveness of the penalty, they do not directly validate this underlying assumption.
  3. The observation subspace is estimated using synthetic corruptions at a fixed severity level, which may not adequately represent real-world variations such as viewpoint changes, occlusion, and background shifts. Additionally, the experiments mainly consider fixed data partitions and probe realizations, leaving the robustness of the method across different acquisition conditions and experimental configurations insufficiently explored.
  4. The covariance adjustment derived from the positive spectrum of the penalized scatter matrix is largely heuristic. The paper lacks a principled explanation of why this particular construction should improve unknown-class rejection, or under what conditions the proposed between-class protection may become ineffective or even detrimental.
Questions for authors
  1. Can the authors clarify the fundamental methodological distinction between PROBE and existing approaches based on nuisance-invariant subspaces, discriminative scatter analysis, and covariance-based open-set scoring? What is the key new principle beyond combining these components?
  2. Is there empirical evidence that directions separating known-class means are also informative for distinguishing known classes from unseen, semantically similar classes? Could the authors directly analyze the projections of unknown-class feature differences onto the protected and unprotected subspaces?
  3. How sensitive is the method to the choice and sampling of corruption operators? In particular, does the improvement persist under genuine acquisition shifts involving viewpoint changes, background variation, or occlusion?
  4. Could the authors provide stronger evidence of the practical significance of the relatively small gains over the strongest baselines on several SSB benchmarks, including uncertainty estimates for paired method comparisons and additional probe realizations or data partitions?
4 comments
AuthorsOpenReview ↗

Official Response to Reviewer stRp (Part 1/2)

Thank you for the careful and insightful review. Your questions led to the new empirical and theoretical analyses below, which strengthen the paper and which we will incorporate into the revised manuscript.

W1/Q1. What is the new principle beyond combining components? The principle is this: a post-hoc Gaussian rejector should add tolerance mainly along directions in which known classes move under observation changes, and penalize the overlap of that tolerance with directions that separate classes. None of the three families you list separates these two roles:

  • Covariance-based scoring, such as Mahalanobis (Lee et al., 2018) and Mahalanobis++ (Müller & Hein, 2025), estimates one within-class covariance. That covariance mixes nuisance variation with semantic variation, so it cannot tell which directions are safe to widen. Adding M2 on top of shared and class-specific Ledoit–Wolf covariances improves AUROC on all six datasets (Table 7 of the paper), so M2 adds directional information beyond the covariance. In new experiments, PROBE also beats Mahalanobis++ on all six datasets, with AUROC gains from 0.92 [0.47, 1.35] to 5.00 [4.60, 5.41] (paired bootstrap 95% CIs in brackets).
  • Discriminative-scatter methods, such as LDA-style scoring and WDiscOOD (Chen et al., 2023), score features in the class-separating subspace and its residual. PROBE uses that subspace in the opposite role, as a constraint on where tolerance may go. PROBE beats WDiscOOD on all six datasets, with gains from 0.73 [0.57, 0.90] to 15.00 [14.34, 15.64].
  • Nuisance-invariant representation learning retrains the encoder to remove nuisance directions. PROBE keeps the encoder and its predictions unchanged, and it uses nuisance directions as tolerance in the rejection score rather than removing them.

The fixed-spectrum control (Table 5 of the paper) tests this principle in isolation. It keeps eigenvalues, trace, and log-determinant identical and changes only the subspace directions. The protected directions win on all six datasets.

W2/Q2. Are known-class-separating directions informative for unknowns? Projection analysis. Yes. We ran the projection analysis you asked for, at two levels.

Sample level. For a test sample, let r=z−μy^r=z-\mu_{\hat y} and define its occupancy of a subspace UU as ρ(U)=∥Ur∥2/∥r∥2\rho(U)=\Vert Ur\Vert^2/\Vert r\Vert^2. The selectivity ratio ρHard unknown/ρknown\rho_{\text{Hard unknown}}/\rho_{\text{known}} measures how much tolerance a subspace gives unknown residuals relative to known ones; below 1 is the desired behavior.

Table R1: Selectivity ratio.

AircraftCarsCUBDTSVisioDECT
UpU_p (PROBE)0.8970.9200.8820.8050.920
U0U_0 (no protection, β=0\beta=0)1.0291.0891.0341.0201.248
Random subspace1.0000.9940.9991.0000.994

Without protection, the observation subspace favors unknown residuals (ratio > 1). With between-class protection, the ratio drops to 0.81–0.92 on every dataset. Protection therefore moves tolerance away from where unknowns lie.

Class level. For each unknown class, let δ\delta be the displacement from its mean to the nearest known-class mean, and let γ\gamma be the fraction of ∥δ∥2\Vert\delta\Vert^2 that lies in the span of SBS_B.

Table R2: Where unknown-class displacements lie.

AircraftCarsCUBDTSVisioDECT
γ\gamma0.9790.9770.9840.9420.951
Chance (random subspace, same dim.)0.0240.0470.0480.0020.002
Share of ∥δ∥2\Vert\delta\Vert^2 in U0U_00.9960.6600.7660.9880.822
Share of ∥δ∥2\Vert\delta\Vert^2 in UpU_p0.4600.3720.3380.4230.518
  • 94–98% of each unknown-class displacement lies in the span of known-class scatter, against 0.2–4.8% by chance. This validates the proxy directly, including on the three fine-grained SSB datasets.
  • Without protection, the observation subspace covers 66–100% of these displacements; protection cuts this to 34–52%.

The analysis covers five datasets; ICG-Drone-3c has two known classes per fold, so its span(SBS_B) is one-dimensional. We will state the proxy explicitly as a premise in the revised manuscript; W4 (Part 2) gives the exact condition under which it would fail.

AuthorsOpenReview ↗

Official Response to Reviewer stRp (Part 2/2)

W3/Q3. Sensitivity to corruption choice; real acquisition shifts. Operator choice. Using all 19 operators gives the best AUROC on four of six datasets and is within 0.4 points of the best single family on the other two. The best family varies by dataset (weather on Aircraft, CUB, and DTS; digital on Cars; blur on VisioDECT; noise on ICG-Drone-3c). Uniform sampling is therefore robust and avoids per-dataset tuning.

Severity. Severity 1 is best on five of six datasets, and higher severity lowers AUROC (Aircraft 84.00 → 82.53, DTS 84.61 → 79.77 at severity 5). Stronger corruptions change predictions more often (Aircraft label preservation falls from 95.7% to 40.1%) and overlap more with between-class directions. Severity 1 keeps probes on the observation side, as intended.

Real acquisition shifts. On the UAV videos, we fit M2 using crops of the same target in consecutive frames. These pairs capture genuine changes in viewpoint, scale, background, and illumination.

Table R3: AUROC (%) with different probe sources.

M1 only (paper Table 4)Real pairsReal + corruption pairsCorruption pairs (paper)
DTS78.6282.4284.1484.61
VisioDECT92.0392.6193.6393.61
ICG-Drone-3c73.8973.3774.6775.88
  • With real frame pairs alone, M2 improves over M1 by 3.80 points on DTS and 0.58 on VisioDECT. The mechanism therefore also works with genuine acquisition changes.
  • Combining real and corruption pairs improves over M1 on all three datasets.
  • Corruption probes alone match or beat real pairs, so synthetic probes suffice without video data.

Q4. Practical significance and uncertainty of the SSB margins. All post-hoc methods share one frozen classifier, so we computed paired bootstrap 95% CIs (2,000 test-set resamples) of PROBE's gain over the strongest other post-hoc method on each dataset.

Table R4: AUROC gain of PROBE (points) [paired bootstrap 95% CI].

Strongest post-hocPROBE gain
AircraftMLS3.18 [2.61, 3.72]
CarsKNN0.46 [0.20, 0.72]
CUBMSP0.61 [0.14, 1.04]
DTSKNN3.40 [3.08, 3.71]
VisioDECTViM0.67 [0.54, 0.80]
ICG-Drone-3cWDiscOOD1.48 [0.41, 2.57]
  • No interval includes zero, so the gains are statistically significant on all six datasets, including the smaller SSB margins on Cars and CUB.
  • The strongest baseline changes from dataset to dataset (MLS, KNN, MSP, ViM, WDiscOOD), whereas PROBE is best on all six. In practice, one method works reliably across benchmarks without selecting a detector per dataset.
  • SSB uses its standard benchmark splits. The UAV protocols use leave-one-class-out evaluation over 7, 6, and 3 folds, so every drone model serves as the unknown class once. For probes, uniform sampling over all 19 operators (W3) removes the dependence on any single operator subset.

W4. A principled explanation, and when protection fails. We derived the following proposition for the appendix. Let a known class be N(μ,s2I)\mathcal N(\mu,s^2I), an unknown class N(μ+δ,s2I)\mathcal N(\mu+\delta,s^2I), and the scoring covariance Σ=s2I+auu⊤\Sigma=s^2I+a uu^\top with ∥u∥=1\Vert u\Vert=1. Because the trace and log-determinant terms cancel, the expected energy gap is

Eunk[E]−Eknown[E]=δ⊤Σ−1δ=∥δ∥2s2−a(u⊤δ)2s2(s2+a).\mathbb E_{unk}[E]-\mathbb E_{known}[E]=\delta^\top\Sigma^{-1}\delta=\frac{\Vert\delta\Vert^2}{s^2}-\frac{a(u^\top\delta)^2}{s^2(s^2+a)}.

  • The gap shrinks only through (u⊤δ)2(u^\top\delta)^2: added tolerance costs separability exactly to the extent that it overlaps the unknown's displacement δ\delta. Tolerance orthogonal to δ\delta costs nothing, while it can still reduce false rejections of corrupted known samples.
  • When δ\delta lies in the span of between-class scatter, the β\beta-penalty limits this overlap. The γ\gamma analysis in W2 (Table R2) shows that this condition holds on all five datasets tested (94–98%).
  • Failure condition. If δ\delta is orthogonal to the between-class span but aligned with observation directions, the penalty provides no protection. With γ≥0.94\gamma\ge 0.94 on all measured datasets, this case does not arise in our benchmarks; we will state it in the Limitations paragraph. Tables 12–15 of the paper support the remaining design choices: positive-spectrum clipping, the isotropic tail τ\tau, and the additive form.

Thank you again; we are happy to clarify any remaining points.

I appreciate the authors’ responses. I still have two questions:

  1. The real acquisition-shift experiments use consecutive frames of the same UAV targets. In addition, the VisioDECT dataset is largely split into adjacent temporal blocks from the same recordings, with no gap between splits. This raises concerns about temporal correlation and recording-specific bias. Can the authors evaluate the method on test data from entirely separate recordings to verify that the improvements generalize beyond the recordings used for training and validation?
  2. The paired bootstrap appears to resample individual test samples, although UAV crops from the same recording may be highly correlated. Could the authors perform a recording-level or sequence-level bootstrap to check whether the reported confidence intervals and statistical significance still hold?
AuthorsOpenReview ↗

Follow-up Response: Separate-Recording Evaluation and Recording/Sequence-Level Bootstrap

Thank you for following up. We evaluated the frozen models without retraining or refitting, and quantified uncertainty using synchronized recording-level and temporal-block bootstrap analyses.

1. Generalization to separate recordings

We evaluated the frozen Sunny-trained ResNet-50 classifiers and Sunny-fitted detectors from Table 2 on the official Cloudy and Evening recordings, disjoint from training, validation, and detector fitting. All 6,768 Cloudy and 6,940 Evening crops (13,708 crops; 12 recordings; six classes) were evaluated under the same six-fold leave-one-class-out protocol. Combined metrics pool both subsets within each fold.

Table F1. Means across folds and three seeds; AUROC / OSCR (%), except for shared known-class accuracy (Acc).

MethodCloudyEveningCloudy + Evening
(Shared-CE Acc)(70.93)(80.67)(75.83)
MSP62.45 / 48.0868.01 / 56.9165.00 / 52.70
MLS58.58 / 44.3763.50 / 51.6860.91 / 48.36
Energy58.08 / 43.8763.04 / 51.1660.43 / 47.85
KNN66.57 / 52.2274.87 / 62.9970.63 / 57.97
Mahalanobis61.94 / 48.7073.59 / 61.8167.69 / 55.50
ViM60.87 / 48.0972.87 / 61.1066.88 / 54.88
Mahalanobis++66.20 / 51.9375.73 / 64.2970.72 / 58.30
WDiscOOD61.07 / 47.5672.09 / 60.1166.55 / 54.17
PROBE68.41 / 54.6077.13 / 65.7972.62 / 60.35
M1 only (α=0\alpha=0)64.61 / 50.2572.30 / 60.4068.56 / 55.86
PROBE (Sunny real pairs)68.67 / 54.1976.55 / 65.0572.25 / 59.70

Default PROBE exceeds all eight competing post-hoc baselines in AUROC and OSCR on all three test sets. Its combined gains over Mahalanobis++ are 1.90 AUROC and 2.05 OSCR points; the AUROC gain over ViM is 5.74 points. Relative to M1 (68.56% combined AUROC), PROBE using Sunny real pairs achieves 72.25% (+3.69 points), and PROBE using Sunny corruption probes achieves 72.62% (+4.06 points). Both variants exceed all eight baselines on both metrics in all three test sets. Frame pairs are used for detector fitting on Sunny training data; inference scores individual crops.

2. Recording-level and temporal-block paired bootstrap

We used whole recordings and contiguous blocks defined once within each complete physical recording and mapped to each fold. Each draw assigns one Poisson(1) weight per unique recording or block, shared across all methods, folds, and three seeds. If any fold has zero total known or unknown weight, we discard the entire draw and resample until obtaining 2,000 valid draws. We compute paired AUROC differences within each seed and fold and average them as in Table 2.

Table F2. Synchronized cluster-bootstrap 95% intervals for AUROC gains (percentage points). K/U: mean known/unknown clusters per fold; total: distinct clusters across folds.

BenchmarkUnit (avg. K/U; total)vs. strongest competing baselinevs. Mahalanobis++vs. WDiscOOD
DTSRecording (7.7/4.1; 29)KNN: +3.40 [1.93, 6.92]+5.00 [3.46, 10.74]+15.00 [11.39, 26.74]
DTS20-crop block (28.3/20.9; 146)KNN: +3.40 [1.39, 5.74]+5.00 [2.07, 8.54]+15.00 [9.14, 21.43]
VisioDECT Sunny10-crop block (117.5/115.5; 693)ViM: +0.67 [0.22, 1.16]+2.04 [1.44, 2.63]+0.73 [0.14, 1.37]
VisioDECT Sunny20-crop block (60.0/57.8; 347)ViM: +0.67 [0.04, 1.28]+2.04 [1.18, 2.85]+0.73 [-0.14, 1.59]
VisioDECT Cloudy+Eve.20-crop block (575.8/115.2; 691)Mah++: +1.90 [1.43, 2.36]+1.90 [1.43, 2.36]+6.07 [5.33, 6.87]
VisioDECT Cloudy+Eve.Recording (10.0/2.0; 12)Mah++: +1.90 [-0.37, 4.31]+1.90 [-0.37, 4.31]+6.07 [2.33, 10.63]
ICG-Drone-3c10-crop block (13.0/35.3; 84)WDisc: +1.48 [-0.88, 3.77]+2.84 [0.06, 5.50]+1.48 [-0.88, 3.77]
ICG-Drone-3cRecording (4.3/11.0; 26)WDisc: +1.48 [-2.63, 5.62]+2.84 [-0.71, 7.39]+1.48 [-2.63, 5.62]
  • DTS: Recording-level and 20-crop intervals exclude zero for all three reported comparisons.
  • Sunny: The 10-crop intervals exclude zero for all three comparisons. The 20-crop intervals exclude zero against ViM and Mahalanobis++ and include zero against WDiscOOD.
  • Cloudy+Evening: The 20-crop intervals exclude zero against Mahalanobis++, WDiscOOD, KNN (+1.99 [1.50, 2.49]), and ViM (+5.74 [5.08, 6.40]). Recording-level intervals exclude zero against WDiscOOD and ViM (+5.74 [2.73, 9.19]) and include zero against Mahalanobis++ and KNN (+1.99 [-0.57, 4.43]).
  • ICG-Drone-3c: The 10-crop interval excludes zero against Mahalanobis++ and includes zero against WDiscOOD. Both recording-level intervals include zero.

We hope Tables F1-F2 thoroughly address your remaining questions. Please let us know if further details would help; otherwise, we hope this new evidence can be taken into account in your final assessment. Thank you again for your insightful feedback.

Reviewer_AuRmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a fine-grained open-set recognition method using frozen classifiers. It combines class-specific isotropic Gaussian energies on unit-normalized features with an additive covariance adjustment estimated from paired clean and mildly corrupted training images. A between-class scatter penalty discourages increased tolerance along directions that distinguish known-class means. The method is fitted in closed form using only known-class training images and their corrupted versions, while keeping the pretrained classifier and its class predictions unchanged. Experiments on three Semantic Shift Benchmark (SSB) datasets and three UAV protocols show competitive performance, with additional evaluations across backbones.

Critical strengths

As a post-hoc method, PROBE can be added to an existing classifier without retraining it or changing its class predictions. It uses only known-class training data and requires a single forward pass at test time, making it relatively straightforward to deploy.

Critical weaknesses
  1. The paper overlooks closely related work, including Mahalanobis++ [1], which normalizes features before Mahalanobis scoring, and WDiscOOD [2], which uses discriminative subspaces for post-hoc detection. Their absence makes it harder to assess the novelty and practical advantage.
  2. Although the paper focuses on fine-grained recognition, PROBE is a general post-hoc detector. Evaluating it on generalized OOD detection benchmarks (such as OpenOOD [3]) would help determine whether its gains extend beyond the selected datasets and enable broader comparisons with existing methods.

[1] Mahalanobis++: Improving OOD Detection via Feature Normalization. ICML 2025. [2] WDiscOOD: Out-of-Distribution Detection via Whitened Linear Discriminant Analysis. ICCV 2023. [3] OpenOOD: Benchmarking Generalized Out-of-Distribution Detection. NeurIPS 2022 Datasets and Benchmarks Track.

Questions for authors
  1. Can you evaluate the proposed method on standard OpenOOD benchmarks? If not, are there any limitations that prevent PROBE from being applied to these benchmarks?
  2. Can you compare the proposed method with closely related methods such as Mahalanobis++ and WDiscOOD?
  3. Was generative AI used to create the illustrations in Figure 2? Could you also show actual augmented training images alongside their clean counterparts to clarify what the corruption looks like in practice?
6 comments
AuthorsOpenReview ↗

Official Response to Reviewer AuRm (Part 1/2)

Thank you for your constructive review and for pointing us to Mahalanobis++, WDiscOOD, and OpenOOD. To address your concerns, we ran all three comparisons; our responses are detailed below.

W1/Q2. Comparison with Mahalanobis++ and WDiscOOD. Mahalanobis++ (Müller & Hein, 2025) and WDiscOOD (Chen et al., 2023) are evaluated on exactly the same frozen classifiers as all other post-hoc baselines. Mahalanobis++ follows its original specification: ℓ2\ell_2-normalized features with a shared covariance.

Table R1: Hard AUROC (SSB) and AUROC (UAV), %. ICG = ICG-Drone-3c.

AircraftCarsCUBDTSVisioDECTICG
Mahalanobis++79.2983.2676.7179.6191.5773.04
WDiscOOD74.3580.5568.6869.6192.8874.40
PROBE84.0086.2177.6384.6193.6175.88

Table R2: Hard OSCR (SSB) and OSCR (UAV), %.

AircraftCarsCUBDTSVisioDECTICG
Mahalanobis++75.5582.1770.3876.8289.0666.71
WDiscOOD70.8879.3862.0067.2089.6468.40
PROBE78.7084.9470.6581.0990.5168.97

PROBE outperforms both methods on every dataset, metric, and backbone tested:

  • Statistical significance. Because all post-hoc methods share one classifier, the comparison is paired. PROBE's AUROC gain over Mahalanobis++ in Table R1 ranges from 0.92 [0.47, 1.35] (CUB) to 5.00 [4.60, 5.41] (DTS), and its gain over WDiscOOD from 0.73 [0.57, 0.90] (VisioDECT) to 15.00 [14.34, 15.64] (DTS). Brackets are paired bootstrap 95% CIs (2,000 test-set resamples); no interval includes zero.
  • SSB Easy split. PROBE achieves 89.95 / 93.43 / 86.09 AUROC on Aircraft / Cars / CUB, compared to 88.06 / 91.51 / 83.94 for Mahalanobis++ and 84.28 / 87.83 / 76.41 for WDiscOOD. Easy OSCR follows the same ordering: 83.62 / 91.62 / 76.85 for PROBE, 83.10 / 90.06 / 75.82 for Mahalanobis++, and 79.65 / 86.42 / 68.28 for WDiscOOD.
  • Other backbones (DenseNet-161, ConvNeXt-T, ViT-B/16 on the three SSB datasets). With both methods added, PROBE remains the best post-hoc method in all nine settings.

How PROBE differs. Mahalanobis++ addresses feature-norm variation; PROBE's M1 also handles it via unit normalization. Our additional standard-OSR runs show how much this matters. On the VGG32 classifiers of the Vaze et al. (2022) splits, unknown samples have markedly smaller feature norms, so raw-feature Mahalanobis (Lee et al., 2018) drops to 43–62 AUROC (e.g., 43.19 on CIFAR-10). The same distance on ℓ2\ell_2-normalized features reaches 75–95 (85.64 on CIFAR-10).

Beyond normalization, PROBE's core contribution is M2, which determines along which directions extra tolerance is added. Table 7 of our paper isolates this effect. Adding M2 on top of a shared Ledoit–Wolf covariance on normalized features, a close variant of Mahalanobis++, improves AUROC on all six datasets: e.g., 79.94 → 80.91 on Aircraft, 81.49 → 82.53 on DTS, and 70.53 → 72.31 on ICG. M2 thus contributes directional information that normalization and covariance estimation do not provide.

WDiscOOD scores features in a discriminative (whitened LDA) subspace and its residual. PROBE uses between-class scatter in the opposite role: as a penalty that keeps observation tolerance out of class-separating directions.

We will discuss both works in Sec. 2 and add both baselines to Tables 1–3 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer AuRm (Part 2/2)

W2/Q1. Evaluation on OpenOOD. Thank you for this helpful suggestion. PROBE applies directly to OpenOOD: as a post-hoc detector, it needs only known-class training images and their mildly corrupted copies. We ran PROBE on OpenOOD v1.5 (Zhang et al., 2023) with the official ResNet-18 checkpoints and the hyperparameters from our paper (α=0.3,β=0.5,q=64\alpha=0.3,\beta=0.5,q=64), without any retuning.

Table R3: OpenOOD v1.5 AUROC (%), mean over the three official checkpoints. C-10 / C-100 / IN-200 = CIFAR-10 / CIFAR-100 / ImageNet-200.

C-10 nearC-10 farC-100 nearC-100 farIN-200 nearIN-200 farAvg.
MSP88.2390.6180.7077.8983.3490.1385.15
Energy87.8991.1681.5580.5882.5090.8685.76
MLS87.8291.0481.6580.3682.9091.1185.81
KNN90.9693.4181.1481.5581.0089.5586.27
ViM88.9294.1075.0680.6278.6891.2684.77
Mahalanobis++88.4194.0574.1682.5981.7794.4985.91
WDiscOOD88.0692.9861.5870.3262.5275.1775.11
PROBE90.8193.2681.3682.4482.6593.7487.38
  • PROBE achieves the highest mean AUROC (87.38), ahead of KNN (86.27) and Mahalanobis++ (85.91).
  • It is the most consistent method. In every setting it is within 0.84 points of the top score, whereas every other method trails the best by at least 3.38 points in some setting.
  • On near-OOD, the setting closest to the fine-grained case PROBE targets, PROBE has the highest average (84.94). It improves over Mahalanobis++ by 2.40, 7.20, and 0.88 points on CIFAR-10, CIFAR-100, and ImageNet-200.

Q3. Figure 2 and actual augmented images. Thank you for asking. We drew Figure 2 by hand in PowerPoint, without generative AI, to convey the design of the method. The generic geometric elements (e.g., spheres and blocks) come from publicly available graphical resources. As noted in our AI-use statement, generative AI was used only for literature retrieval.

The revised manuscript will include side-by-side clean and severity-1 training images for all four corruption families. Severity-1 operators introduce very mild perturbations, such as slight sensor noise, minimal blur, faint weather effects, or minor contrast changes. They leave fine-grained attributes intact: the frozen classifier keeps its prediction on 94.9–99.7% of corrupted training images.

We hope these results and clarifications address your concerns, and we welcome any further feedback.

Dear Authors,

Thank you for your response.

You mentioned that “the revised manuscript will include side-by-side clean and severity-1 training images for all four corruption families.” Could you upload the revised manuscript and highlight the changes in a different color? This would help me better understand your revision.

Best regards, Reviewer

Reviewer_zUvmOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper improves the post-hoc rejection energy for frozen classifiers that combines class-specific isotropic energies on unit-normalized features with a between-class-protected additive covariance adjustment for fine-grained OSR and evaluates on the commonly used testbench.

Critical strengths
  1. The paper is complete in writing. The overall structure of the paper follows standard conventions (Introduction, Related Work, Method, Experiments), making the sections easy to locate.

  2. The paper includes comparisons against several established baseline models.

Critical weaknesses

The paper is in general not easy to follow.

  1. The paper is hard to read. Some abbreviations are not introduced before being used, such as {(zjr,zja,yjr)}j=1P\{(z_j^r, z_j^a, y_j^r)\}_{j=1}^P in line 157 and d in equation 2, although d for distance is conventionally used in Gaussian energy methods, and citations are missing, such as lines 035- 039 and 044- 049.

  2. The main problem to address is not well demonstrated, or the problem formulation is unclear. What is the connection between observation-induced variance and fine-grained OSR.

  3. If there are comparisons on the normal OSR testbench, that would be better.

  4. The paper assumes that the fine-grained open sets lie close to the edge of the class support of the known classes. However, the validation and supporting citations for this assumption are missing.

  5. The reasons for adapting the applied corruption are unknown.

  6. Missing ablation: the effects of the severity of the corruptions on the performance of PROBE. Would increased severity harm the performance?

  7. Missing important baselines and related work [1, 2, 3]

[1] WDiscOOD: Out-of-Distribution Detection via Whitened Linear Discriminant Analysis, ICCV'23

[2] ViM: Out-of-Distribution With Virtual-Logit Matching, CVPR'22

[3] Tangent Prop - A formalism for specifying selected invariances in an adaptive network, NIPS'91

Questions for authors
  1. See 3 in Weaknesses: Is PROBE designed only for fine-grained OSR or for OSR in general?

  2. See 4 in Weaknesses: Why does the assumption hold? Is it still suitable for normal OSR?

  3. See 5 and 6 in Weakness: why are these corruptions and severity level are applied? Would increased severity harm the performance?

  4. See 7 in Weakness: Please differentiate PROBE from the related work listed. Otherwise, this method looks incremental.

6 comments
AuthorsOpenReview ↗

Official Response to Reviewer zUvm (Part 1/2)

We sincerely thank the reviewer for the thorough review and constructive comments. Below we address each concern with additional experiments and clarifications for the revision.

W1/W2. Undefined symbols and missing citations. Thank you for noting these presentation issues. We will revise the text as follows:

  • L157: The superscripts in (zjr,zja)(z_j^{r}, z_j^{a}) will be defined at first use (rr: clean reference image, aa: its corrupted copy).
  • L155 (Eq. 2): Sd−1\mathbb{S}^{d-1} will be defined at first use as the unit sphere in Rd\mathbb{R}^d, where dd denotes the feature dimension.
  • L035–039 and L044–049: We will add a citation for each claim. Fine-grained unknowns lying close to known classes and differing in subtle local attributes: Vaze et al. (2022) and Lang et al. (2024). Feature shifts under changing imaging conditions: Hendrycks & Dietterich (2019). Feature norm varying with prediction confidence: Park et al. (2023). Norm variation degrading Gaussian scoring: Müller & Hein (2025). Within-class covariance scoring: Lee et al. (2018).

W3/Q1. How does observation-induced variation relate to fine-grained OSR? Is PROBE only for fine-grained OSR? Thank you for this important question. In fine-grained OSR, an unseen bird species can lie closer in feature space to a known species than a blurred or noisy photo of that known species does. A post-hoc Gaussian rejector must accept the blurred known photo. If it gains that tolerance by widening the class isotropically, or along directions that separate classes, it also accepts the nearby unknown. PROBE addresses this trade-off. It learns from paired clean/corrupted images which directions observation changes move a feature along, and adds variance mainly along those directions while penalizing their overlap with class-separating directions. We will state this connection explicitly in the first paragraph of the Introduction.

PROBE targets fine-grained OSR and applies unchanged to general OSR. We evaluated it on two general benchmarks.

Table R1: Standard OSR (splits of Vaze et al., 2022; VGG32 CE classifier; AUROC %, mean over 5 splits).

CIFAR-10CIFAR+10CIFAR+50TinyImageNet
MSP90.1295.1693.3474.93
Energy92.6597.4795.7376.81
KNN89.0495.7593.1075.92
Mahalanobis++85.6494.8591.4274.64
ViM87.7292.8991.5874.60
WDiscOOD88.3795.9293.1074.56
PROBE90.4595.1792.7577.03
  • PROBE achieves the best AUROC (77.03) and OSCR (64.36, vs. 64.12 for MLS) on TinyImageNet, the benchmark with the most classes and the closest unknowns. On CIFAR, where known and unknown classes are coarse categories (e.g., animals vs. vehicles), logit-based scores such as Energy perform best.
  • On OpenOOD v1.5 (Zhang et al., 2023; official ResNet-18 checkpoints for CIFAR-10, CIFAR-100, and ImageNet-200), PROBE achieves the highest mean AUROC over the six near/far-OOD settings: 87.38, versus 86.27 for KNN, 85.91 for Mahalanobis++, and 85.76 for Energy. It is also the only method within 1 point of the best score in every setting.

PROBE thus transfers to general OSR/OOD benchmarks without modification, and its largest gains appear when unknowns are semantically close to known classes.

W4/Q2. Do fine-grained unknowns lie near the edge of known-class support? Does this hold in normal OSR? We validated this assumption directly. For each test sample, we measured its normalized squared distance to the predicted class mean, ∥z−μy^∥2/(dσy^2)\Vert z-\mu_{\hat y}\Vert^2/(d\sigma^2_{\hat y}).

Table R2: Medians, and the 90th percentile (p90) of known samples.

Known medianKnown p90Hard unknown medianEasy unknown median
Aircraft1.917.066.327.09
Cars1.744.735.597.66
CUB2.394.974.234.60
VisioDECT (UAV)1.006.3214.01 (held-out class)–
  • On all three SSB datasets, Hard unknowns lie closer to known classes than Easy unknowns.
  • On Aircraft and CUB, the median Hard unknown falls inside the known p90, and on Cars it sits just beyond it. Fine-grained unknowns therefore lie at the boundary of known-class support. This matches the design of SSB, where Hard unknowns are selected for high semantic similarity to known classes (Vaze et al., 2022; Lang et al., 2024).
AuthorsOpenReview ↗

Official Response to Reviewer zUvm (Part 2/2)

W4/Q2 (continued).

  • In normal OSR, coarse unknowns lie much farther away. On VisioDECT, the held-out median (14.01) is more than twice the known p90, which makes the tolerance–rejection trade-off milder. PROBE is still the best method on VisioDECT (93.61 AUROC) and remains competitive on CIFAR and OpenOOD (W3). The property thus explains where directional protection helps most, and PROBE stays strong where it holds less.
  • We also verified the mechanism directly, by measuring how much of each test residual falls into the subspace where tolerance is added. Relative to known samples, PROBE's protected subspace gives Hard-unknown residuals only 0.81–0.92 times as much tolerance. An unprotected subspace (β=0\beta=0) gives them 1.02–1.25 times as much.

W5/W6/Q3. Why are these corruptions and this severity level applied? Would increased severity harm performance?

  • Why ImageNet-C operators: They are a standardized, reproducible set of common acquisition degradations (noise, blur, weather, digital processing; Hendrycks & Dietterich, 2019). These are the nuisance factors a rejector should tolerate without treating them as a change of class.
  • Why severity 1: At severity 1, the frozen classifier keeps its prediction on 94.9–99.7% of corrupted training images, so the probe pairs reflect changes of observation rather than changes of class. We will add this explanation to Sec. 3.1.

Severity ablation. We evaluated PROBE with probes at severities 1, 3, and 5.

Table R3: AUROC (%) by probe severity.

sev 1sev 3sev 5
Aircraft84.0083.7182.53
Cars86.2185.6385.33
CUB77.6376.8076.44
DTS84.6182.4979.77
VisioDECT93.6192.1691.83

Higher severity reduces performance, for the reason above. From severity 1 to 5, label preservation drops markedly (95.7% → 40.1% on Aircraft, 94.9% → 65.1% on DTS), and the overlap between probe displacements and between-class directions grows (0.13 → 0.61 on VisioDECT). Severe corruptions begin to destroy class-discriminative attributes, so the probes no longer isolate observation directions. On ICG-Drone-3c, the smallest dataset, all severities stay within 0.8 points of the default.

Operator choice. Sampling all 19 operators uniformly gives the best AUROC on four of six datasets and is within 0.4 points of the best single family on the other two. Because the best single family varies across datasets (weather on Aircraft, CUB, and DTS; digital on Cars), uniform sampling avoids per-dataset tuning.

W7/Q4. Differentiation and comparison with WDiscOOD, ViM, and Tangent Prop. We newly evaluated WDiscOOD (Chen et al., 2023) on all datasets and backbones. Below we report it alongside ViM (Wang et al., 2022a), which was already a baseline in Tables 1–3.

Table R4: Hard AUROC (SSB) and AUROC (UAV), %. ICG = ICG-Drone-3c.

AircraftCarsCUBDTSVisioDECTICG
ViM66.9479.3069.2376.0092.9471.31
WDiscOOD74.3580.5568.6869.6192.8874.40
PROBE84.0086.2177.6384.6193.6175.88
  • ViM was already discussed in Sec. 2 (L107–108). PROBE outperforms it on all six datasets, on TinyImageNet (77.03 vs. 74.60, Table R1), and in mean OpenOOD AUROC (87.38 vs. 84.77).
  • WDiscOOD: PROBE outperforms it on all six datasets, with paired AUROC gains of 5.66–9.65 points on SSB and up to 15.00 on DTS (no paired bootstrap 95% CI includes zero), and in all nine additional backbone settings. Conceptual difference: WDiscOOD scores features inside a whitened LDA subspace and its residual. PROBE uses between-class scatter the other way round: as a penalty that keeps added tolerance out of class-separating directions, while placing tolerance along observation directions learned from probe pairs.
  • Tangent Prop (Simard et al., 1991) regularizes network training so that outputs become invariant along transformation tangents. PROBE is post-hoc and leaves the network unchanged. It uses observation directions only inside the rejection energy, as added variance, and it penalizes overlap with between-class scatter, which has no counterpart in Tangent Prop.

We will incorporate all of these results and discussions into the revised manuscript. We hope this addresses your concerns and welcome any further feedback.

Could you please also report the change in close-set accuracy for the severity ablation?

Comments on the paper

2 comments

Dear reviewers,

Thank you for your reviews and constructive feedback.

The authors have submitted their rebuttal. Please take a moment to review it and consider whether your concerns have been adequately addressed. It would be helpful if you could briefly indicate:

  • Whether the rebuttal addresses your main concerns;

  • Whether any significant issues remain unresolved;

  • Whether your overall recommendation has changed.

If appropriate, please also update your review and recommendation in the system.

Thank you for your time and valuable contributions.

Best regards,

AC

Area_Chair_HMqbOpenReview ↗

AC update

Dear authors,

Thank you for your time and contributions to the discussion.

I understand that you may be concerned if some reviewers have not provided follow-up comments or explicitly acknowledged how their concerns have been addressed in the rebuttal and revised manuscript. Please be assured that the AC will carefully consider all available information, including the original reviews, author rebuttal, revised manuscript, and discussion comments, as part of the final assessment.

The absence of further reviewer comments does not mean that the authors’ responses and revisions have been overlooked. I will carefully examine the issues raised, assess the authors’ clarifications and revisions, and take the discussion into account when forming my recommendation. I will also consult with the reviewers as appropriate to ensure that the relevant issues are fully considered.

Thank you again for your constructive engagement and contributions to the review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.