- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces an OOD detection method for image classifiers by analyzing the spectral projection profile of the penultimate feature of a network. It connects the spectral projection profile to well-behavedness showing that more well-behaved models have a more balanced profile while less well-behaved models tend to concentrate its ID representations around the leading right singular directions. OOD representations, on the other hand, have the same concentration around the leading directions even in well-behaved classifiers and this gap correlates with how well-behaved a classifier is. Additionally, it shows that shape bias is represented in the trailing spectral directions. Using these observations, the paper proposes a combination of two methods: OMA, a logit-based method that focuses on leading spectral directions and SHU that tries to detect differences in the trailing spectral directions. Together the method OMASHU++ and its training statistics assisted variant OMASHU beat the competition on four selected well-behaved image classifiers on ImageNet-1k using the OpenOOD framework.
Critical strengths
- The paper establishes a connection between the final layers spectral projection profile and the well-behavedness of a model empirically by showing a strong correlation between the projection gap of the leading directions and well-behavedness.
- It empirically shows that the shape bias of a model is more represented in the leading spectral directions.
- The proposed method outperforms the competition on the evaluated models.
Critical weaknesses
- Selecting models by largest projection gap is generally unfair for the competition as one can expect the method to perform well there. The authors argue to use well-behavedness as selection metric which still introduces a bias, but is a broader concept and not directly connected to the method. That the spectral projection gap correlates with well-behavedness is a novel insight. Arguments why this is the case could improve the connection and make the selection more defendable.
- The method is only evaluated on four well-behaved classifiers.
- The method is only evaluated on ImageNet-1K and not on other ID datasets.
- SHU is fitting the storyline well, especially that shape bias is represented in the spectral projection profile, but is not adding much to the performance.
Questions for authors
-
I am not sure how shuffling affects the ID and OOD profile. I suspect non-shuffled ID samples do not exhibit a drop in activation in the tail directions. Maybe the Figure 2 could be extended by also showing the original ID profile.
-
Do you have any theoretical argument or intuition why the spectral projection profile could be linked to well-behavedness or shape bias?
2 comments
Dear Reviewer uEop, we thank you again for the thoughtful feedback. We have uploaded the revised paper, with changes marked in blue. In brief, regarding your main concerns: the correlation with well-behavedness now rests on 21 models; experiments with additional backbones and dataset; unshuffled ID and OOD profiles in Figure 2. Our response further clarifies the contribution of SHU and the intuition behind the link to well-behavedness. We would be grateful to hear whether this addresses your concerns, or whether anything remains that we can clarify.
Response to Reviewer uEop
W2 and W3 are addressed in G3 and Q1 in G2 of our general response.
W1: Justifying the selection of well-behaved models
We select models by well-behavedness, not by projection gap. Our motivation stems from the observation that OOD detection methods built on older backbones may not generalize to modern ones. The well-behavedness benchmark (Beyond Accuracy, Hesse et al. 2026) already establishes that well-behaved backbones, which tend to be recent, outperform less well-behaved ones across multiple dimensions, yielding them more desirable in practice. Indeed, well-behavedness and the projection gap strongly correlate with each model’s publication date, with Pearson correlations of 0.75 and 0.77, respectively. Our analysis highlights the limitations of existing OOD detection methods in well-behaved models, and our method overcomes them by leveraging insights from spectral projection profile analysis. We give further intuition for the correlation between well-behavedness and the projection gap in Q2.
W4: Contribution of SHU
We kindly note that unlike stated, SHU substantially improves OOD detection performance. For example, in Table 1 of the main paper adding SHU to OMA increases AUC by 2.15 pp and 2.25 pp and decreases FPR by 5.42 pp and 4.75 pp (similar improvements hold for SHU++).
Q2: Intuition for the link to well-behavedness and shape bias
Intuitively, the leading singular directions of the linear head capture the most dominant, widely shared feature variations, such as local texture statistics, that appear in both ID and OOD images. Well-behaved models are trained with modern recipes (e.g., strong augmentation, Mixup/CutMix, label smoothing) that reduce texture shortcuts, so their ID-class evidence also relies on global cues, spreading ID features beyond the leading directions, consistent with their robustness. OOD images, in contrast, lack the global cues learned for the ID classes and remain concentrated on the leading, texture-related directions, which yields the projection gap.