- Rating
- 1
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper introduces Mixed-Granularity Open-Set Recognition (MG-OSR), a setting in which known labels may occupy different levels of an underlying category hierarchy while satisfying an antichain constraint, and the hierarchy itself is hidden from the learner. The paper argues that this regime changes the meaning of “unknown”: an unseen subclass of a coarse-grained known class should still be accepted, whereas a neighboring category outside a fine-grained known class should be rejected. The main contributions are the MG-OSR formulation, a controlled benchmark construction, an analysis of asymmetric failure modes and cross-granularity interference, and a granularity-aware evaluation protocol intended to expose failures that aggregate OSR metrics such as AUROC and OSCR can obscure. Validation is framed as a benchmark-based diagnostic and re-evaluation of existing conventional and VLM-based OSR methods rather than as a new recognition method.
Critical strengths
Strength 1: The conceptual reframing of “unknown” relative to the scope of known labels is clear and potentially impactful. The paper makes a simple but important point: being unseen during training does not uniquely determine whether a sample should be treated as unknown; the answer depends on the granularity at which the known label space is defined. This is consequential because it identifies a blind spot in standard OSR evaluation and gives the community a precise vocabulary for discussing scope-dependent knownness.
Critical weaknesses
Weakness 1: The MG-OSR setting and benchmark construction have limited ecological validity, and the central conclusions may be partly artifacts of the chosen taxonomy and split protocol. The definition requires known labels to form an antichain and keeps the hierarchy hidden from the learner, while benchmark construction uses the taxonomy to decide which descendants are unseen known variants and which neighbors are near unknowns. Real open-set deployments often involve known labels with ancestor–descendant relations, partially available semantics, or ambiguous scope boundaries; the paper does not establish how often the antichain regime occurs or how conclusions change when it is relaxed. This matters because the paper’s main contribution is the claim that mixed granularity is an overlooked and important regime. If the regime is narrowly constructed and highly dependent on hidden taxonomic choices, the broader impact is weakened.
Weakness 2: The paper is primarily a benchmark and analysis paper, but the evidence chain from the proposed diagnostics to the strong claims about representation learning and evaluation is not fully established. The paper explicitly states that it focuses on exposing and evaluating mixed-granularity failures rather than solving them, which limits the methodological contribution. Moreover, the cross-granularity interference analysis reports reduced gradient coherence among sibling categories, but does not establish a causal link between this coherence measure and downstream OSR performance, nor does it rule out alternative explanations such as label noise, class imbalance, or data distribution shifts induced by the split construction. The H-Score is also threshold-dependent and still an aggregate summary. For a paper whose main claims are about failure modes and evaluation mismatch, this makes the evidence more suggestive than decisive, which is a substantive rejection concern.
Weakness 3: The manuscript as provided has missing or corrupted figures.
Questions for authors
The provided manuscript has a corrupted figure region: Page 5 is largely repeated “1”s, Figure 1 is missing, and Figures 2 and 3 are unreadable with truncated captions. Can you confirm whether this is a submission or compilation error and provide a complete, high-resolution manuscript? Without readable Figures 2 and 3, the central claims about cross-granularity interference and evaluation mismatch cannot be assessed.
How common is the antichain constraint in realistic open-set deployments? If known labels can have ancestor–descendant relations, how should MG-OSR be extended, and would the main failure modes reported here persist?
The benchmark construction uses the hidden taxonomy to define unseen known variants and near unknowns. How sensitive are the conclusions to the choice of taxonomic branches, coarse/fine composition, and hierarchy depth? Can you report results under alternative split constructions?
H-Score depends on the TPR95 operating point and the harmonic mean of coarse-variant acceptance and near-unknown rejection. How stable are the method rankings under other thresholds or alternative aggregation choices?
The cross-granularity interference analysis reports reduced gradient coherence among sibling categories. Is this coherence measure correlated with per-branch OSR performance, and can an intervention that changes coherence produce a corresponding change in recognition behavior?
Given that no recognition method is proposed, how do the authors position the methodological contribution of this paper relative to benchmark and analysis papers? This directly affects how the contribution should be weighed.