acceptodds
Withdrawn from ICLR 2027

GeoHOC: Geometric Hierarchical Out-of-Distribution Classification via Nested Angular Caps

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Many visual recognition tasks involve both hierarchical class structures and unknown classes in an open world. Hierarchical classification focuses on fine-grained recognition along the class hierarchy, while out-of-distribution recognition concerns when an input no longer belongs to the existing classes. Unifying the two requires jointly modeling the semantic direction and acceptance extent of each class, so that the model can both localize an appropriate semantic branch and determine the proper prediction depth. Based on this perspective, we propose GeoHOC, a geometric framework that jointly models semantic directions and acceptance extents. In a normalized vision–language space, each category node is represented by an angular cap with a learnable center and radius. A joint loss encourages correct branch selection and parent–child containment, while promoting image acceptance by correct caps and rejection by incorrect sibling caps. During inference, center similarity selects a candidate child branch, while a calibrated cap-based acceptance rule determines whether to descend or stop at the current node. We further introduce a known-only calibration scheme to select the stopping offset of this acceptance rule. Experiments on three benchmark datasets demonstrate that GeoHOC achieves the best overall performance among the compared methods, effectively balancing known class recognition and ancestor recovery for unseen classes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
hierarchical classification, open-set recognition, out-of-distribution classification, vision-language models, geometric representation learning

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_rzgnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper proposes GeoHOC, a geometric framework for hierarchical out-of-distribution classification that aims to recognize known classes while assigning unseen classes to their appropriate semantic ancestors. The method represents each category as an angular cap in a normalized vision-language embedding space, with a learnable center and radius controlling semantic direction and acceptance extent. Training combines hierarchical path discrimination, parent-child containment regularization, and cap-based discrimination to encourage correct branch selection and rejection of incorrect sibling classes. During inference, the model recursively selects the most similar child based on its center and uses a calibrated cap acceptance threshold to determine whether to continue descending or stop at the current node. The authors also introduce a known-only calibration strategy that simulates unseen classes by masking known child categories when selecting the stopping threshold. Experiments on CIFAR-100, CUB-200, and Stanford Cars demonstrate improvements in the combined known-class recognition and novel-parent recovery metric, with additional ablations examining the contribution of the geometric objectives and learned parameters.
Critical strengths
  • The paper presents a clear and reasonably well-motivated geometric formulation for hierarchical out-of-distribution classification. Representing each class with a learnable angular center and radius provides an intuitive way to connect semantic classification with prediction granularity. The separation between branch localization and cap-based acceptance is also practical and easy to understand.
  • The empirical evaluation is reasonably thorough, covering three datasets with multiple folds or seeds, comparisons against recent hierarchical novelty detection methods, and ablations of the main components. I particularly appreciate the inclusion of matched-backbone cosine baselines, geometric containment diagnostics, and recognition-fallback trade-off analysis, which provide useful evidence beyond the primary accuracy metric.
Critical weaknesses
  • While the overall results are positive, the empirical advantage is somewhat less convincing than the main presentation suggests. GeoHOC achieves the highest MixBAcc on all three datasets, but the improvements over the strongest baselines are only 1.37 and 0.76 percentage points on CUB-200 and Stanford Cars, respectively. Moreover, ProHOC achieves higher novel-parent accuracy on both datasets, and GeoHOC does not achieve the best AUROC on any of the three benchmarks. The main comparisons also involve different backbone architectures, making it difficult to isolate the contribution of the proposed geometry from the underlying visual representations. Although the matched-backbone ablations help, the improvement over the depth-wise cosine baseline on CUB-200 is relatively modest.

  • The geometric formulation is intuitive, but its practical relationship to the claimed nested acceptance regions is not entirely clear. The containment constraint is enforced through a soft penalty rather than guaranteed, and Table 4 shows that approximately 39.4% of parent-child edges still violate the sufficient nesting condition on CUB-200. The calibrated stopping offset further changes the effective acceptance boundaries, so the extent to which the final decision regions preserve the intended hierarchical geometry remains unclear. In addition, all three evaluations focus on unseen fine-grained classes with retained direct parents, leaving the behavior under more general hierarchical novelty scenarios largely untested.

Questions for authors
  • The improvements over the strongest baselines on CUB-200 and Stanford Cars are relatively small, while the novel-parent accuracy is lower than ProHOC on both datasets. Could the authors provide a clearer comparison of the known-versus-novel trade-off against the strongest existing methods, preferably under matched backbone settings? Since the method combines OpenCLIP adaptation with learned node-specific acceptance boundaries, stronger evidence isolating the benefit of the geometric formulation would improve my assessment.
  • Table 4 reports that approximately 39.4% of parent-child edges still violate the sufficient containment condition, even after applying the nesting loss. How often do the learned caps actually fail to satisfy geometric containment, and how does this change after applying the calibrated stopping offset? It would also be useful to understand whether enforcing stricter nesting improves or hurts the final classification results.
  • The reported AUROC is lower than the strongest baseline on all three datasets. Appendix D.2 indicates that the novelty score combines cap-based risk with frozen-CLIP similarity risk. Could the authors report the AUROC of these two components separately and clarify how much the learned cap geometry contributes to out-of-distribution discrimination?
  • The experiments consider unseen leaf classes whose direct parent remains in the existing hierarchy. How does GeoHOC behave when unseen classes appear at higher hierarchical levels, or when no meaningful existing ancestor is available? Since the method is framed as hierarchical out-of-distribution classification, evidence beyond the held-out leaf setting would help establish the scope of its applicability.
Reviewer_qd6SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoHOC addresses hierarchical open-set classification, where an image from an unseen class should be assigned to its most specific known ancestor. The method represents each modeled class as an angular cap in a normalized vision–language embedding space. Each cap has a learnable center and radius, and training combines hierarchical classification, parent–child containment, and cap-based acceptance and rejection losses.

At inference, the model selects a child using center similarity and then decides whether to descend using that child’s cap boundary. A stopping offset is calibrated using known-class images, combining ordinary classification with a leave-one-out procedure that masks the correct class and asks the model to return its parent.

Experiments on CIFAR-100, CUB-200, and Stanford Cars show the highest mean MixBAcc among the compared methods. The paper also includes component ablations, matched-backbone threshold baselines, and geometric diagnostics.

Critical strengths
  1. The method gives a clear interpretation to hierarchical acceptance decisions. Representing a node through a semantic direction and an acceptance radius is a sensible design. The inference rule is simple, and the calibration procedure avoids using held-out novel classes. The authors also ensure that masked-child calibration retains at least one competing sibling, so parent recovery cannot be achieved merely because no candidate remains.
  2. The evaluation includes useful controls beyond comparisons with existing methods. The matched-backbone cosine classifiers use the same representation, adaptation, and calibration objective, which helps separate the contribution of geometry from that of OpenCLIP. The ablations show that removing nesting or cap discrimination reduces mean MixBAcc on CUB. The recognition–fallback curves provide further evidence that nesting can improve parent recovery at matched levels of known-class accuracy over part of the operating range.
Critical weaknesses
  1. The advantage over strong alternatives is modest and not yet sufficiently isolated. GeoHOC improves mean MixBAcc over the strongest baseline by 6.47 points on CIFAR-100, but only 1.37 points on CUB and 0.76 points on Cars. On the latter two datasets, ProHOC achieves higher novel-parent accuracy, and GeoHOC does not achieve the best AUROC on any dataset. These results support a favorable operating-point trade-off, but not a broad improvement in novelty recognition. The matched-backbone comparison is helpful, but the depth-wise cosine baseline already reaches 47.52 MixBAcc on CUB versus 48.69 for GeoHOC. Since an angular radius corresponds to a node-specific cosine threshold, a classifier with independently learned node thresholds would be a particularly relevant control. The no-nesting ablation partially addresses this issue, but the remaining gains would be more convincing with paired per-run comparisons and matched-backbone results on all three datasets.

  2. The connection between nested geometry and the final acceptance rule remains incomplete. The nesting loss regularizes the raw caps, whereas inference uses the shifted threshold (z^\top e_v \geq \cos\phi_v-\delta). This shift changes each cap’s effective angular radius and does not automatically preserve parent–child containment. The reported geometric diagnostics concern only the raw caps. Even there, the full model violates the margin-based containment condition on 39.41% of edges under the reported depth-balanced statistic. This does not invalidate the predictor, and the appendix correctly notes that violating the sufficient condition does not necessarily imply failure of actual containment. Nevertheless, the paper needs a clearer account of how much hierarchical nesting remains after calibration and whether it explains the observed benefit. The current evaluation also holds out leaves whose direct parents remain known, leaving the broader claims about variable prediction depth and root-level rejection largely untested.

Questions for authors
  1. Can you provide paired per-run differences against ProHOC and the matched-backbone depth-wise threshold baseline? Are the smaller MixBAcc improvements consistent across folds or seeds?
  2. How does GeoHOC compare with a matched-backbone model using independently learned node-specific cosine thresholds? This would help distinguish the benefit of hierarchical geometric coupling from the flexibility of class-specific acceptance boundaries.
  3. What are the containment statistics for the effective acceptance regions after applying the selected stopping offset? Do these statistics relate to parent-recovery performance more closely than the raw-cap diagnostics?
  4. All five CIFAR-100 runs select (\delta=0.50), the upper boundary of the search grid. Does expanding the grid change the selected offset or the reported results? Also, how well does masked-known-class parent recovery predict actual novel-parent recovery across runs?
  5. Do you have results for held-out subtrees or inputs with no modeled ancestor? These cases would clarify whether the method’s benefits extend beyond stopping at the immediate parent of an unseen leaf.
Reviewer_C6LcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeoHOC, a geometric framework for hierarchical out-of-distribution classification. The method represents each category as an angular cap with a learnable center and radius in a shared vision-language embedding space. Experiments on three benchmarks demonstrate improvements in the overall balanced accuracy.

Critical strengths
  1. The paper addresses hierarchical OOD classification by jointly considering fine-grained recognition of known classes and ancestor-level prediction for unseen classes, which is practically relevant to open-world recognition.
  2. Representing categories as angular caps provides a unified way to model semantic directions and class acceptance regions.
  3. The paper evaluates GeoHOC on three benchmarks and includes ablation studies.
Critical weaknesses
  1. The key motivation of GeoHOC is to represent hierarchical class relationships through nested angular caps. However, Equation (12) only softly encourages the containment condition in Equation (8). Table 4 shows that 39.41% of parent-child edges still violate the margin-based condition, with a mean positive excess of 0.04019 radians, exceeding the nesting margin of 0.02 radians. More importantly, the stopping offset introduced in Equation (16) changes the effective angular radius to ϕveff=arccos⁡(cos⁡ϕv−δ)\phi_v^{\mathrm{eff}}=\arccos(\cos\phi_v-\delta). Due to this nonlinear transformation, the original containment condition does not guarantee that the effective acceptance regions remain nested after calibration, even when Equation (8) is fully satisfied. This creates a potential inconsistency between the geometric constraints used during training and the decision boundaries employed during inference.
  2. The method selects the child with the highest center similarity and checks only its cap acceptance condition. However, since different children have different angular radii, the most similar child does not necessarily have the highest acceptance margin. If the selected child is rejected, the model immediately returns the parent without considering other potentially valid siblings. So it suggests that more analysis on the potentially valid siblings.
  3. Insufficient validation of known-only calibration for unseen classes. The masked-child strategy simulates unknown classes by removing known categories during calibration. But, these samples still originate from classes seen during training and may not reflect the feature distributions of genuinely unseen classes.
Questions for authors
  1. See Weaknesss.
  2. Appendix D.3 uses substantially different hierarchical classification loss weights across datasets λhier=1.0\lambda_{\mathrm{hier}}=1.0 for CIFAR-100 and (0.07) for CUB-200 and Stanford Cars. What motivates this difference, and how sensitive is the performance to these settings?
  3. In Appendix A.1, internal-node centers are initialized from text embeddings but subsequently optimized as independent parameters. How is semantic alignment with the updated text representations maintained during training?

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

Related papers

Discussion (0)

Sign in to comment.