acceptodds
Under review as a conference paper at ICLR 2027

OpenMorphoQuery: From Visual Morphology Prompts to Target-Grounded Queries for Unseen Industrial Defect Localization

Abstract

Under a category-disjoint open-set setting, localizing unseen industrial defects from scarce visual morphology prompts requires transferring fine-grained morphology revealed by exemplars to target images. Such transfer characterizes morphology-defined openness, whose central challenge stems from a reference-frame mismatch in morphological representation. Within prompt RoIs, discriminative morphology is encoded relative to material texture, product context, and annotation boundaries rather than as an independently transferable category attribute. Conventional aggregation preserves this sample-dependent anchoring, yielding prototypes that characterize the intended morphology but cannot establish reliable correspondence with its target-image manifestation. We propose OpenMorphoQuery (OMQ), a morphology-guided framework that addresses this mismatch through progressive reference-frame transformation. The Normalized Morphology Dictionary Encoder re-expresses context-bound RoI features relative to shared morphology bases through normalized residual coding, establishing a transferable morphology frame. Perturbation-Consistent Prototype Experts retain relations stable under appearance, geometry, and boundary perturbations, further stabilizing this frame against incidental variations. Prototype-Transport Query Initialization then transports the stabilized prototypes into a target-conditioned frame through optimal-transport alignment, yielding spatially grounded decoder queries. Experiments on a unified four-dataset benchmark of over 160,000 images establish state-of-the-art aggregate performance in multi-granular morphology grounding among visual-prompt methods. Ablations further quantify the incremental contribution of each component.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
industrial defect localization, open-set recognition, morphology representation, few-shot learning, optimal transport, query-based detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_29BkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OpenMorphoQuery (OMQ), which detects and segments industrial defects unseen during training using a few region-annotated visual exemples. It extends DINOv with three modules that reduce feature entanglement, refine defect prototypes, and initialize queries using target-image features. Across 4 datasets with 147 seen and 52 unseen categories, OMQ achieves box and mask AP scores of 28.5 and 21.5, outperforming DINOv by 6.5 and 5.3%p, respectively. Ablation studies and response alignment analyses further support its effectiveness.

Critical strengths
  • The paper addresses a practically relevant problem through an intuitive interface. The reported performance supports the effectiveness of its approach to addressing feature entanglement in existing methods.
  • The method delivers substantial improvements over its base model, DINOv, in both box and mask AP. Performance improves as the proposed modules are added cumulatively.
  • In the prompt-scaling experiments, the method consistently outperforms the baseline across 1~5-shot settings. OMQ also achieves better results in the response alignment analysis.
  • The method extends DINOv with a small parameter overhead of approximately 200k additional module parameters.
Critical weaknesses
  • Evidence for the central mechanism is indirect. Table 5 quantitatively demonstrates better alignment between model responses and actual defect regions. However, the analysis does not sufficiently distinguish whether these improvements arise from mitigating the claimed morphology reference-frame mismatch or from more general benefits of feature aggregation, normalization, or improved query initialization.
  • Although the main hyperparameter values are disclosed in Table A2, their selection rationale and the sensitivity of performance to these settings are insufficiently analyzed. In particular, it is difficult to assess the effects of the NMDE dictionary size and temperature, the PCPE weights, and the PTQI weight settings.
  • Validation of the robustness and scope of the improvements is limited. The paper does not report variability across training seeds or resampled support sets, making it difficult to determine whether the observed gains persist under these changes. Moreover, on GC10-DET, box and mask AP are 7.5 and 4.3 %p lower than those of DINOv, respectively. Further analysis is needed to explain this degradation and clarify the method's limitations.
  • Comparative analysis of training and inference efficiency is insufficient. Computational efficiency is an important consideration when deploying the method in industrial inspection settings. The paper does not compare training time, inference latency, or memory usage against DINOv under matched conditions, making it difficult to assess the trade-off between the performance gains and the computational cost of the additional modules.
Questions for authors

Clarifying the following points would help assess the methodological contributions and interpret the experimental results.

  1. Mechanism validation: Could you provide controlled experiments to establish whether the gains arise specifically from reducing morphology reference-frame mismatch in exemplar features, beyond the benefits of general feature processing or improved query initialization?
  2. Robustness and failure conditions: Do the improvements persist across multiple training seeds and resampled support sets? What accounts for the lower performance relative to DINOv on GC10-DET?
  3. Hyperparameters: What validation procedure was used to select each module's hyperparameters? Could you report how performance varies with the main hyperparameter settings?
  4. Computational efficiency: How much do training time, inference latency, and memory usage increase relative to DINOv under identical hardware and input conditions? These measurements would help assess the performance gains relative to the additional computational cost.
Reviewer_5mWrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies unseen industrial defect localization using visual morphology prompts, focusing on transferring fine-grained defect morphology from scarce exemplars to target images. The authors characterize the central transfer challenge as a morphology reference-frame mismatch, in which prompt-RoI morphology is encoded relative to material texture, product context, and annotation boundaries. They propose OpenMorphoQuery (OMQ)—comprising a Normalized Morphology Dictionary Encoder (NMDE), Perturbation-Consistent Prototype Experts (PCPE), and Prototype-Transport Query Initialization (PTQI)—and evaluate it through performance comparisons, component ablations, and prompt-target alignment analyses on a four-dataset category-disjoint industrial benchmark.

Critical strengths
  1. The paper addresses unseen industrial defect localization using visual exemplars to specify fine-grained defect morphology.

  2. The method combines morphology encoding, prototype stabilization, and target-grounded query initialization. The ablation results in Table 3 show that adding each module improves both box and mask AP.

  3. The evaluation covers four industrial datasets and includes detailed category-level results and discussion of cases where OMQ underperforms DINOv.

Critical weaknesses
  1. Baseline comparability is unclear. Lines 285–291 and 320–323 do not specify baseline-specific fine-tuning or frozen components, and the UniSpector reimplementation is insufficiently validated. Table 3 supports incremental module gains, but does not resolve possible adaptation and implementation differences in the cross-method comparisons.

  2. The reference-frame interpretation lacks sufficient mathematical justification. Lines 174–179 describe a shared morphology frame, but Eq. (2), zc,k,m=∑nAc,k,n,m(xc,k,n−bˉm)∑nAc,k,n,m+ϵ,z_{c,k,m}=\frac{\sum_n A_{c,k,n,m}(x_{c,k,n}-\bar b_m)}{\sum_n A_{c,k,n,m}+\epsilon}, does not explain why residual encoding separates morphology from context. Likewise, minimizing 1−cos⁡(pc,kr,pc,kr,e)1-\cos(p^r_{c,k},p^{r,e}_{c,k}) in Eq. (7) encourages perturbation consistency, but the paper does not establish that the chosen perturbations preserve defect-defining morphology.

  3. The evidence does not isolate the proposed mechanism. With 88% defect-free evaluation images (Lines 279–283), AP reflects both localization and false-positive suppression. Table 5 measures output alignment rather than dependence on support context. Better target-side discrimination could improve both, leaving the reference-frame interpretation in Lines 437–446 insufficiently distinguished from alternative explanations.

Questions for authors
  1. For the methods compared in Table 1, please specify the training data, fine-tuning schedule, and which components were frozen or updated for each baseline. How did the authors verify that their UniSpector reimplementation reproduces the original method’s performance?

  2. What evidence shows that the residual encoding in Eq. (2), zc,k,m=∑nAc,k,n,m(xc,k,n−bˉm)∑nAc,k,n,m+ϵ,z_{c,k,m}=\frac{\sum_n A_{c,k,n,m}(x_{c,k,n}-\bar b_m)}{\sum_n A_{c,k,n,m}+\epsilon}, reduces sensitivity to support context while preserving sensitivity to defect morphology? Likewise, minimizing 1−cos⁡(pc,kr,pc,kr,e)1-\cos(p^r_{c,k},p^{r,e}_{c,k}) in Eq. (7) encourages agreement under the perturbations in Eq. (3), but how do the authors distinguish removal of nuisance variation from loss of discriminative morphology? Can they compare representation changes under context-only and morphology-changing modifications, before and after NMDE and PCPE?

  3. Given that 88% of evaluation images are defect-free (Lines 279–283), can the authors evaluate the same trained models with and without defect-free queries and separately report false positives on defect-free images at a matched recall level? Do OMQ’s gains persist when evaluation is restricted to defective images, and how does including defect-free images change the performance gap?

Reviewer_Mh46OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper proposes OpenMorphoQuery (OMQ) for localizing unseen industrial defects specified by visual morphology exemplars rather than fixed category labels.
  2. The Normalized Morphology Dictionary Encoder, Perturbation-Consistent Prototype Experts, and Prototype-Transport Query Initialization progressively transform prompt features into target-grounded queries.
  3. The benchmark combines four public industrial datasets and evaluates 147 seen categories and 52 unseen categories using a category-disjoint protocol.
  4. OMQ is evaluated with one, three, and five visual exemplars per category, alongside visual-prompt and open-vocabulary baselines.
  5. OMQ achieves box/mask AP of 28.5%/21.5% on the pooled benchmark and improves over DINOv overall, although the source-specific results show a substantial regression on GC10-DET.
Critical strengths
  1. The paper identifies a concrete problem in visual prompting: defect morphology can be entangled with the product context and annotation boundaries of the support images.
  2. The category-disjoint benchmark covers multiple industrial datasets and includes 52 unseen test categories.
  3. The visual-prompt comparisons use the same fixed support exemplars, making the main comparisons easier to interpret.
  4. The ablations quantify the separate contributions of morphology encoding, target grounding, and prototype calibration.
  5. The response-alignment analysis provides additional evidence beyond aggregate detection and segmentation metrics.
Critical weaknesses
  1. OMQ performs substantially worse than DINOv on GC10-DET, where box/mask AP decreases from 13.1%/8.9% to 5.6%/4.6%, despite improvements on the other three source-specific datasets.
  2. The pooled benchmark results do not fully convey this source-specific regression, making the overall improvement less uniform than the headline results suggest.
  3. The method depends on annotated visual support regions for unseen categories, and the associated annotation effort is not evaluated.
  4. The robustness of the method to inaccurate support boxes, different exemplar selections, and noisy morphology cues remains unclear.
  5. The evaluation uses four public datasets, so generalization to independently collected industrial data and new acquisition pipelines remains unestablished.
Questions for authors
  1. What causes the performance drop on GC10-DET, and which defect categories or visual characteristics account for it?
  2. Can the authors report macro-averaged results across datasets and categories alongside the pooled benchmark results?
  3. How sensitive is OMQ to the choice of support exemplars and to errors in the annotated support regions?
  4. Can the method be evaluated on an independently collected industrial dataset that was not used to construct the benchmark?
  5. How do inference latency and memory usage scale with the number of prompted categories, target-image tokens, and localization queries?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.