acceptodds
Under review as a conference paper at ICLR 2027

ZAC : Zero-shot Anomaly-Aware Counting for Industrial Components

Abstract

Counting industrial components is essential in real-world manufacturing systems, and what matters is not only how many parts there are but how many are usable, since production lines carry defective items alongside intact ones. As manufacturing lines continually introduce new components with distinct defect types, no single model can be trained on every class it will encounter. To address this, we introduce Zero-shot Anomaly-Aware Counting, a task that counts objects while distinguishing their defect states. Traditional object counting models focus on quantity and ignore quality, while industrial anomaly detection models evaluate quality but struggle with dense multi-object scenes. Multimodal large language models can be queried for either, but their performance drops substantially once a query asks which of the objects they have found are defective. These failures share a cause, since existing benchmarks do not combine instance-level defect-state supervision with class-disjoint evaluation on unseen component classes. Therefore, we construct Real-ZAC, a benchmark of real industrial components in dense multi-instance settings. Every instance is annotated with an intact or defective label, and test classes are strictly withheld from training to evaluate zero-shot capabilities. Furthermore, we propose a baseline method that utilizes the multi-instance structure of Real-ZAC. The model suppresses visual patterns shared across instances and emphasizes subtle defect-related deviations, reducing reliance on class-specific appearance and improving generalization to unseen component classes. The code and dataset will be available soon.

Primary area:
datasets and benchmarks
Keywords:
Anomaly-Aware Counting, Object Counting, Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bxttOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a very clear approach to anomaly-aware counting in industrial scenes. The method first localizes multiple ROIs in an image, then compares local features across different instances in the same scene to highlight regions that deviate from the others, allowing the model to distinguish intact from defective instances and count them accordingly. The overall idea is straightforward and easy to follow, and the authors also build a relatively large Real-ZAC dataset and conduct extensive experiments.

Critical strengths
  1. A major strength is the real industrial setting and the large dataset, which gives the experiments solid support.
  2. The core idea also makes a lot of sense for this problem, and the experiments show that it works well.
Critical weaknesses

A key weakness is the limited novelty. Recent work such as MuSc and Odd-One-Out already uses very similar ideas based on mutual matching or cross-instance comparison, so this paper currently feels more like applying an effective existing idea to a new industrial setting.

What I would really like to see is stronger evidence that the method learns a genuinely different representation for this task. For example, showing that the learned features become more defect-sensitive and less dominated by object identity, rather than just improving performance through a known comparison strategy.

Questions for authors
  1. Could the authors clarify the key algorithmic novelty of Pinpoint beyond prior comparative anomaly-detection methods such as MuSc and Odd-One-Out? In particular, what new capability or representation is introduced here, rather than adapting an existing cross-instance comparison idea to a new setting?

  2. Can the authors provide stronger evidence that Pinpoint actually learns a different representation for defect state? For example, it would be useful to show whether the learned features become more defect-sensitive and less dominated by object identity after adding the comparative descriptor and within-image centering.

  3. For the ObjEmbed baseline in Table 2, does it use exactly the same CountGD++ proposals and the same training setup as Pinpoint? Since the paper shows that proposal quality has a large effect on performance, this detail is important for understanding how much of the gain comes from the proposed representation itself.

8 comments

We thank the reviewer for the thoughtful review and appreciate the positive assessment of the method’s practicality.


Q1.

While MuSc and Odd-One-Out both compare instances, Pinpoint differs in what the comparison is used for. It turns a within-image comparison into a per-instance intact/defective decision for a queried class on unseen classes, without any threshold. Neither prior method provides this capability, which rests on three design choices described below. We have cited and discussed Odd-One-Out in Section 2 of the revision.

(1) Training-free comparison. Odd-One-Out improves generalization through DINOv2 feature distillation, but its matching features, the attention over matched voxels and the anomaly decision are still learned on seen categories. Pinpoint instead computes the comparison by mutual scoring on frozen DINOv3 features without any learned parameters, so the comparison itself is not fitted to the training classes.

(2) Class-independent reference point. MuSc outputs raw distances whose scale varies with the class and the reference images, so it can rank instances but cannot classify them with a fixed rule. Within-image centering (Eq. 4) expresses each instance relative to its own image, so that zero means "typical for this image" in every class. This step yields the largest single gain in our ablation.

(3) Language-based decision. The comparison serves as a cue rather than as the decision itself. It is injected into a region token that also carries semantic information, and the intact/defective query margin (Eq. 8) makes the final decision. This restricts the decision to the queried class and provides a fixed rule that transfers to unseen classes.

The task and its benchmark. These design choices are driven by the requirements of a new task, which we introduce together with its benchmark. Zero-shot anomaly-aware counting asks a model to count the intact and defective instances of a queried component in a single real image, on classes never seen during training. No existing setting combines single-image input, per-instance defect state, counting output and a class-disjoint split. Real-ZAC supports this task with real industrial components, instance-level intact/defective labels, class-disjoint splits and a protocol that evaluates counting error and per-instance state jointly, whereas the ToysAD-8K and PartsAD-15K datasets of Odd-One-Out are synthetic. Pinpoint shows that the task is solvable and provides a reference point for future methods on this benchmark.

Q2.

SettingMAE ↓RMSE ↓F1F_1 ↑
Representation components (validation-selected threshold)
ObjEmbed fine-tuned (original recipe)2.733.70.802
+ LoRA with focal loss2.403.49.834
+ DINOv3 comparative descriptor2.623.86.843
+ within-image centering1.973.04.872
+ magnitude restoration (Pinpoint)1.933.02.876
Pinpoint, inference rule
Validation-selected threshold1.933.02.876
Argmax (t=0t=0, no calibration)2.093.29.872
Plain DINOv3 ROI feature (no mutual scoring)
ROI feature, validation-selected threshold2.513.65.837
+ centering2.173.27.858
ROI feature, argmax (t=0t=0)2.463.73.849
+ centering2.463.69.855

We have extended Table 10 (Table 11 in the revision) with plain DINOv3 ROI-feature rows, so that each component is isolated under identical proposals and training. Within-image centering subtracts the median descriptor of the image. This removes the appearance shared by all instances of the same object and keeps only how each instance deviates from the others. It is the largest single gain in the component chain (MAE 2.62 → 1.97, F1F_1 .843 → .872), larger than adding the comparative descriptor itself. This is consistent with the linear probing in Figure 10, where the probe AUROC increases from .798 to .843 as the comparative descriptor and within-image centering are added.

Q3.

The proposals are identical. ObjEmbed in Table 2 uses the same CountGD++ proposals, NMS and point matching as Pinpoint, and is trained on the same Real-ZAC training split. The only difference is the training recipe, as ObjEmbed is fine-tuned following its original recipe. The gain of Pinpoint over ObjEmbed therefore does not come from the proposals. To avoid confusion, we will state the shared proposals in Section 5.2

Method (Proposals)MAE ↓RMSE ↓F1F_1 ↑
ObjEmbed (CountGD++)2.733.70.802
ObjEmbed (WeDetect)3.265.27.708
Pinpoint (CountGD++)2.093.29.872
Pinpoint (WeDetect)2.474.85.815
Reviewer_C16nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Industrial inspection needs to know how many components are in a frame and how many of them are usable, including for component classes not seen in training. The paper defines zero-shot anomaly-aware counting, where zero-shot means class-disjoint rather than training-free, and makes two contributions. Real-ZAC is a benchmark of 5,485 images over 25 real component classes (21.4 instances per image), with every instance point-annotated as intact or defective (117,124 instances) and five classes held out from training. All-intact and all-defective images are included at 8-12% per class to catch models that rely on odd-one-out comparison (Sec. 3.2). Pinpoint is a baseline that uses class-agnostic proposals, compares instances within the image to suppress their shared appearance, removes class identity by median centering, and reads the residual as a defect cue (Sec. 4). On the five unseen classes it reports F1 0.872 and MAE 2.09, 7.0 F1 points above ObjEmbed-2B (0.802).

Critical strengths

S1. The benchmark is built to expose a specific failure mode. Sec. 3.2 includes all-intact and all-defective images because "a model that judges an instance by comparing it against its neighbors has no reference to rely on in these images". This is a trap for the paper's own method, and Sec. 5.4 reports the all-defective case failing: the reference shifts toward the defective appearance and the defect signal is subtracted away. To my knowledge no existing dataset combines per-instance defect labels with class-disjoint evaluation on real photographs, and this is the main reason for my positive rating.

Critical weaknesses

W1. The headline F1 is dominated by the easy query type. Table 7 shows that the 0.872 F1 pools defective and intact instances whose ground truth splits 4,212 : 17,746 (19% : 81%). Separating defective from intact instances is what distinguishes this task from plain counting, and on the defective query F1 is 0.600 (lv3) and 0.621 (lv4), and 0.244 on one unseen class (resistor, Table 8). The pooled MAE of 2.09 weights the two queries equally, but the defective target is much smaller. With 3.84 defective instances per test image, its MAE of 1.87 is roughly a 49% error. Sec. 1 and Table 2 should report the defective-query numbers alongside the pooled ones.

W2. The method's novelty is weak. Sec. 4.3 distinguishes Pinpoint from MuSc by comparing patches within one image rather than across test images, but Table 14 shows that restricting MuSc to within-image comparison raises its instance-level AUROC from .800 to .857, close to Pinpoint's .858, with no trained parameters (Pinpoint has 12.7M). In Table 10, most of Pinpoint's counting gain comes from LoRA with focal loss on ObjEmbed (+0.032 F1) and the median-centering step (+0.029), while the mutual-scoring descriptor that Sec. 4.3 contrasts with MuSc adds +0.009 and magnitude restoration +0.004. The task framing is also not new. Odd-One-Out (Bhunia et al., CVPR 2025) already infers normality per scene and flags the deviating instance by cross-instance comparison on unseen categories, with its own benchmarks. It is multi-view by design (five views per scene by default) and is not cited. The paper's distinct ground is the single-image, count-producing version of that task, and it should be positioned that way.

Questions for authors

Q1. Results on the all-intact and all-defective subsets are not reported, although Sec. 3.2 and Sec. 5.4 discuss them. Could you report MAE / RMSE / F1 separately for all-intact, all-defective and mixed images of the five unseen classes? At 8-12% per class, the headline 0.872 may come mostly from mixed scenes, where an odd-one-out detector is enough.

Q2. Sec. 4.5 and Table 14 say thresholds (and, for Table 14, the training epoch) are "selected on held-out validation categories", but Table 5 splits the 25 classes into 20 train and 5 test with no validation split. Which categories were used, and were any of the five test classes involved?

Q3. As a benchmark paper, Real-ZAC should document how its 117,124 labels were produced: number of annotators, the intact/defective criteria, inter-annotator agreement, where the components came from, and whether the defects are real production failures or induced. Licence and distribution terms are also missing. Could these be added?

Q4. Real-ZAC arranges instances "without occlusion" (Sec. 3.1), but on a production line parts often overlap or lie in contact. Sec. 5.4 already notes, on MVTec LOCO, that a single proposal can cover a cluster of parts. How do the proposals and the within-image reference behave on Real-ZAC classes when instances partly overlap, for example on a small set of such images?

4 comments

Thank you for the careful review and constructive feedback. We provide our responses below.


Q1.

ModelCompositionImagesObjectsDefectiveMAE ↓RMSE ↓P ↑R ↑F1F_1 ↑
ObjEmbedIntact only5334501.051.51.832.849.841
(+ LoRA, focal loss)Defective only472552554.386.17.241.286.262
Mixed99621,3583,9572.383.56.837.864.850
All1,09621,9584,2122.403.64.829.857.843
PinpointIntact only5334500.591.15.901.919.910
Defective only472552554.916.33.158.188.172
Mixed99621,3583,9572.043.15.867.894.880
All1,09621,9584,2122.093.29.858.886.872

We report MAE, RMSE, precision, recall and F1F_1 separately for all-intact, all-defective and mixed images of the five unseen test classes. ObjEmbed is fine-tuned on the Real-ZAC training split with LoRA and focal loss, following the same recipe as Pinpoint but without the comparative descriptor, and both models use the argmax rule (t=0t=0). As the reviewer notes, mixed images make up most of the test set, and the overall F1F_1 of .872 closely follows their result. We regard this as the intended setting rather than an easy case, because on a production line most components are intact and defective ones are the exception. Within-image comparison alone is also not sufficient. MuSc with within-image comparison reaches an average F1F_1 of .781 over the five folds, and plain DINOv3 ROI features with the same centering reach .855 on the main split, both below Pinpoint. On all-intact images, which correspond to a defect-free tray, the comparative branch reduces false positives. All-defective images account for about 4% of the test set, and accuracy is lower on them, as already discussed in Section 5.4. We will add these results to Appendix C.6.

Q2.

None of the five test classes were involved. The validation set consists of 120 images held out from four randomly chosen training classes, namely fuse, heat shrink tubing, ceramic capacitor and ceramic fuse. These images are not used for training.

The validation set is used for checkpoint or epoch selection, including Table 14 (Table 15 in the revision), and for the quantile cut of the object-level queries (levels 1–2). The component ablation (Table 11) additionally reports rows whose intact/defective threshold is selected on the validation set. All main anomaly-aware counting results come from the defect-state queries (levels 3–4), which use the argmax rule (t=0t=0) and involve no threshold.

Table 5 lists these four classes under Train because they belong to the training split. We have stated the validation set and checkpoint selection in Section 4.5 of the revision.


Q3.

We have added the following details to Appendix A.4, with a pointer in Section 3.2.

  • Defect definition and annotator training. For each component class, we first defined the defects that occur in that component and grouped them into four categories (fracture, deformation, missing part and surface damage; Table 5). Annotators were trained on these class-specific definitions with visual examples before labeling.

  • Defect design. Defects were chosen and produced with four considerations.

    • (i) Realism: each defect reproduces a failure mode that occurs in actual use or handling of that component.
    • (ii) Visibility: every defect is visible from the top-down capture view, so that its state can be judged from a single image.
    • (iii) Identity preservation: a defective instance remains recognizable as its class, so that it is still counted as the queried component.
    • (iv) Diversity: defects vary in severity and location, from subtle to obvious, so that the state cannot be inferred from a single fixed cue.
  • Components and defects. Some components were collected from actual production use and others were purchased commercially. The defects include both real defects and physically induced defects that reproduce the defined failure modes.

  • Annotation. Eight trained annotators labeled every instance with a point and an intact or defective state.

  • Quality control. Each image was checked by two reviewers and then by one final reviewer. Disagreements were resolved by the final reviewer, so every released label reflects a consensus across these stages.

  • License. Real-ZAC will be released under CC BY-NC 4.0 for research use, together with the annotations and the class-disjoint splits.


Q4.

We agree that components on a production line can overlap or lie in contact. As stated in Section 3.1, Real-ZAC deliberately excludes such cases. When part of a component is hidden, whether a defect lies in the occluded region cannot be verified from the image, so its intact or defective label would be ambiguous.

Our goal is to provide a basic benchmark for anomaly-aware counting in which every instance-level label is verifiable, and we therefore did not consider occluded instances.

Extending anomaly-aware counting to overlapping instances, including predicting the state of partly hidden components, is an important direction, and we hope future work will build on Real-ZAC to address it. We will discuss this case in Section 5.4.

Most of my questions are resolved, though the defective-query F1 (0.600/0.621) should appear next to the pooled numbers. I agree with Reviewer bxtt that the benchmark is the main contribution and keep my score of 3.

Thank you for the helpful suggestion and for the continued engagement.

Following your comment, Table 2 now reports the F1 on defective queries alone (F1$_{\text{def}}$) next to the pooled F1 for all compared methods, so that defect-state performance can be read directly alongside the pooled numbers. We also agree that the benchmark is the main contribution of this work, and we have revised the paper accordingly. Pinpoint is now presented consistently as a baseline, and the main text places more emphasis on the construction of Real-ZAC. The changes are marked in red in the revised version.

We sincerely appreciate your constructive feedback throughout the discussion.

Reviewer_1YV3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces zero-shot anomaly-aware counting, where a model must localize and count intact or defective industrial components from categories excluded from task-specific training. The authors construct Real-ZAC, containing 5,485 images across 25 component classes with 117,124 instance-level state annotations. They also propose Pinpoint, which augments ObjEmbed region representations with comparative descriptors extracted using a frozen DINOv3 encoder. The descriptors combine within-image patch matching, anomaly-weighted pooling, median centering, and magnitude scaling, and are injected into region tokens through a learned projection and gate. Training uses LoRA and focal loss.

On the main class-disjoint split, Pinpoint achieves MAE 2.09 and pooled F1 0.872, compared with 2.73 and 0.802 for fine-tuned ObjEmbed. The paper includes five-fold class-disjoint evaluation, component and backbone ablations, and transfer experiments on reannotated multi-instance subsets of VisA, MVTec AD 2, and MVTec LOCO.

Critical strengths
  1. A meaningful problem and a useful benchmark design. Counting usable components requires instance-level state recognition that conventional object-counting and image-level anomaly metrics do not directly evaluate. Real-ZAC combines instance-state annotations, multiple backgrounds, and class-disjoint evaluation at a substantial annotation scale. Its inclusion of both mixed-state and homogeneous-state images creates opportunities to study when comparative recognition succeeds or fails. The external annotations further extend the resource contribution beyond the authors' acquisition setting.

  2. Substantial improvements over a strong starting point. Pinpoint improves pooled F1 from 0.802 to 0.872 and reduces MAE from 2.73 to 2.09 relative to fine-tuned ObjEmbed. The benefit is not restricted to the majority intact class: defective-query F1 increases from 0.427 to 0.621 in Table 6. These results support the practical usefulness of adapting region representations to defect-state recognition.

  3. Evidence beyond a single split or model configuration. Table 9 reports five-fold class-disjoint evaluation, with mean pooled F1 of 0.894 and a fold-to-fold standard deviation of 0.023. Tables 10–13 examine representation components, inference rules, visual and language backbones, and proposal sources. Although the controls could be strengthened, this is a substantive empirical investigation.

  4. External evaluation and informative failure analysis. The comparison on reannotated external datasets uses shared proposals to control localization differences and demonstrates useful transfer of defect-state representations. The authors explicitly discuss thin components, entirely defective images, and mixed-category scenes as limitations. This analysis helps define the scope of the baseline and identifies research challenges that the benchmark can support.

Critical weaknesses

W1. Dependence on within-image references needs clearer characterization

The comparative descriptor measures differences between instances, but such differences need not correspond to defects. In Equation (2), repeated defects may find similar patches in other defective instances and therefore receive low anomaly scores. Median centering can additionally suppress a defective appearance shared by many instances. Section 5.4 explicitly acknowledges failure when every instance is defective, yet the paper does not quantify performance on this subset.

Table 5 already lists 53 all-intact, 47 all-defective, and 996 mixed-state test images. Reporting results separately, with and without the comparative branch, would clarify where the branch helps and where the original ObjEmbed pathway must compensate. This limitation does not negate the aggregate improvements or the benchmark's value, but it constrains claims that the comparative representation captures defectiveness independently of the surrounding instances.

W2. Some uncertainty remains about the source of the representation gains

Table 10 shows that LoRA with focal loss raises F1 from 0.802 to 0.834 before the comparative descriptor is introduced. Adding the DINOv3 descriptor increases F1 to 0.843 but worsens MAE from 2.40 to 2.62; within-image centering then provides the largest subsequent improvement. These results are informative, but do not establish whether mutual scoring is needed beyond access to DINOv3 features and defect-state supervision.

The most useful additional control would be ordinary DINOv3 ROI-feature injection, with and without centering, under identical proposals, supervision, and inference rules. This would distinguish the value of the proposed comparison from the benefit of adding strong visual features. Within-image MuSc also achieves AUROC 0.857 in Table 14, compared with 0.858 for Pinpoint, despite a substantial counting-F1 gap. Separating ranking, proposal recall, and calibration would sharpen the explanation of that gap. The current ablations support the effectiveness of the complete design, but provide a less definitive account of why each operation is necessary.

W3. Several evaluation details need clarification for reproducibility

The paper repeatedly refers to held-out validation categories for threshold or checkpoint selection, but Table 5 assigns all 25 categories to 20 training and 5 test classes. Please specify which training categories are reserved for validation, whether the model is subsequently retrained, and how this procedure is applied in the five-fold experiment. This is a reproducibility concern, not evidence that test labels were used improperly.

External F1max is explicitly obtained by sweeping thresholds on evaluation labels. This is a valid separability measure, but should be distinguished from counting performance under a fixed decision rule. Please specify the thresholds used for Table 4's MAE and RMSE and explain the query, aggregation, or training differences behind the different F1max values in Tables 3 and 4. Figure 10 also includes a “score centering” step and reports query AUROC 0.891, whereas Tables 9 and 14 report 0.858. Its definition and relationship to the final inference procedure need clarification. These differences may have legitimate protocol explanations; the manuscript should make them explicit.

W4. State-specific reporting and dataset documentation would strengthen the resource contribution

The headline F1 of 0.872 pools intact and defective queries, while intact instances account for 80.8% of the main test set. Table 6 reports defective-query F1 of 0.621 and recall of 0.513, and Table 8 reports defective F1 of 0.244 for resistors. The appendix therefore reveals important challenges that deserve greater visibility in the main results, alongside state-macro F1. These remaining errors do not undermine the substantial improvement over the baseline; they also illustrate why the proposed task is worth studying.

Because the dataset is the strongest contribution, the paper should also document defect provenance, annotation verification, the number of distinct physical components, and repeated photography. The five-fold evaluation is valuable, but currently includes only Pinpoint, so it demonstrates stability across splits without establishing that the relative advantage over the strongest baseline persists.

Questions for authors
  1. What are the counting errors and relevant state-specific metrics on the all-intact, all-defective, and mixed-state subsets, with and without the comparative branch? Does the benefit change when several instances share the same defect?

  2. Under identical proposals and inference rules, how does Pinpoint compare with ordinary DINOv3 ROI-feature injection, particularly with the same centering operation but without mutual scoring?

  3. Which categories are used for validation in each split? What exactly is the score-centering operation in Figure 10, and is it used in the final reported model? Please also clarify whether Table 14 uses identical proposals for Pinpoint and the anomaly-detection baselines.

  4. What are the external counting and defective-instance recognition results using thresholds determined entirely from source-domain validation data, or the stated argmax rule? Please explain the protocol differences between Tables 3 and 4 and the thresholds used for their counting metrics.

  5. How were defect labels verified, and were physical components reused across images? If available, matched five-fold results for the strongest baseline would also help establish the stability of the relative improvement.

5 comments

We thank the reviewer for the valuable comments and helpful suggestions. Below we address each concern in detail.


Q1.

ModelCompositionImagesObjectsDefectiveMAE ↓RMSE ↓P ↑R ↑F1F_1 ↑
ObjEmbedIntact only5334501.051.51.832.849.841
(+ LoRA, focal loss)Defective only472552554.386.17.241.286.262
Mixed99621,3583,9572.383.56.837.864.850
All1,09621,9584,2122.403.64.829.857.843
PinpointIntact only5334500.591.15.901.919.910
Defective only472552554.916.33.158.188.172
Mixed99621,3583,9572.043.15.867.894.880
All1,09621,9584,2122.093.29.858.886.872

We agree that results on all-intact and all-defective images are necessary to fully characterize the method, and we now report them separately, with and without the comparative branch. The model without the branch is ObjEmbed fine-tuned on the Real-ZAC training split with LoRA and focal loss, following the same recipe as Pinpoint but without the comparative descriptor, and it uses the same proposals, queries and argmax rule (t=0t=0). In practice, most components on a production line are intact, and defective ones are the exception. Real-ZAC reflects this, as mixed images make up most of the test set. On these images, Pinpoint matches its overall performance, and the comparative branch reduces counting errors and improves F1F_1. On all-intact images, which correspond to a defect-free tray, the comparative branch reduces false positives. The benefit does change when all instances in an image are defective, which is the extreme case of instances sharing a defect. All-defective images are rare, accounting for about 4% of the test set (47 of 1,096 images), and the comparative branch does not help on them. As already discussed in Section 5.4, when a frame contains only defective instances, the within-image reference drifts toward the defective appearance, so the deviation that marks a defect is partly removed. We will add the results to Appendix C.6.

Q2.

SettingMAE ↓RMSE ↓F1F_1 ↑
Representation components (validation-selected threshold)
ObjEmbed fine-tuned (original recipe)2.733.70.802
+ LoRA with focal loss2.403.49.834
+ DINOv3 comparative descriptor2.623.86.843
+ within-image centering1.973.04.872
+ magnitude restoration (Pinpoint)1.933.02.876
Pinpoint, inference rule
Validation-selected threshold1.933.02.876
Argmax (t=0t=0, no calibration)2.093.29.872
Plain DINOv3 ROI feature (no mutual scoring)
ROI feature, validation-selected threshold2.513.65.837
+ centering2.173.27.858
ROI feature, argmax (t=0t=0)2.463.73.849
+ centering2.463.69.855

We trained plain DINOv3 ROI-feature injection, with and without within-image centering, under the same CountGD++ proposals, supervision and gated injection as Pinpoint, and evaluated it under both inference rules. With the same centering but without mutual scoring, plain ROI features are consistently worse than Pinpoint under both rules, and their centering gain largely disappears under the argmax rule. The improvement therefore comes from centering mutual-scoring comparisons, not from the DINOv3 features themselves. We will add these rows to the component ablation (Table 10 in the submitted version, Table 11 in the revision).

Q3.

Validation set The validation set consists of four categories selected at random from the training split, namely fuse, heat shrink tubing, ceramic capacitor and ceramic fuse. They belong to the 20 training classes and are disjoint from the five test categories, while the 120 validation images are held out from the training images. The validation set is used for three purposes. First, we select the checkpoint with the highest instance-level AUROC averaged over the five queries of the four query levels, where level 4 has both a defective and an intact query. Second, we set the quantile cut for the object-level queries (levels 1–2). Third, the component ablation (Table 11) additionally reports rows whose intact/defective threshold is selected on the validation set. All main anomaly-aware counting results come from the defect-state queries (levels 3–4), which use the argmax rule (t=0t=0) and involve no threshold. In the five-fold experiment, the validation classes are drawn only from the training classes of each fold. We have stated the validation set and checkpoint selection in Section 4.5 of the revision.

Evaluation Details on Figure 10 and Tables 9 and 14 The AUROC values in Figure 10 differ from those in Tables 9 and 14 for two reasons. First, Tables 9 and 14 count unmatched background proposals as negatives, whereas Figure 10 evaluates matched objects only. This is necessary for the probing analysis, because the linear probe classifies intact and defective instances and background proposals carry no defect-state label, so both curves in Figure 10 are computed on the same matched objects. For the final model, this difference alone changes the AUROC from .871 (matched objects only, the second-to-last point of Figure 10) to .858 (with background proposals, as in the tables). Second, the last point of Figure 10 is an exploratory variant that is not part of the final model. It adds score centering, which subtracts the mean score of the instances in the same image from each instance’s score, and reaches .891 on matched objects. This score centering is a different operation from the within-image centering of Eq. 4, which subtracts the median descriptor of the image and is part of the final model. To avoid this confusion, we will state these settings explicitly in the caption of Figure 10 and in the accompanying text.

Identical proposals for Pinpoint and the anomaly-detection baselines All methods in Table 14 use the identical CountGD++ proposals, so the differences reflect how each method scores the same regions rather than proposal quality. We will state this explicitly in Appendix E.1 and in the caption of Table 14 (Table 15 in the revision).

Comments on the paper

2 comments

We thank all reviewers for their constructive feedback. We have addressed each reviewer’s concerns in the individual responses below and revised the paper accordingly. All changes in the revised paper are marked in red.

The main concerns and our responses are summarized below.

  • Results by image composition. We report results separately for all-intact, all-defective and mixed images, with and without the comparative branch (Appendix C.6).
  • Novelty and component analysis. We clarify how Pinpoint differs from MuSc and Odd-One-Out and add a plain DINOv3 ROI-feature control under identical proposals and inference rules (Appendix D.1). Odd-One-Out is now discussed in Section 2.
  • Evaluation protocol. We specify the validation set and checkpoint selection (Section 4.5), the shared CountGD++ proposals across methods (Section 5.2) and the cross-dataset setting (Section 5.3). We also report external results under each model’s fixed decision rule, without any threshold tuned on evaluation labels.
  • Stability across splits. We add anomaly-detection baselines to all five class-disjoint folds, and Pinpoint achieves the best F1F_1 in every fold (Appendix C.3).
  • Dataset documentation. We describe the annotation protocol, quality control, component sources and license (Appendix A.4).
AuthorsOpenReview ↗

Updated General Response to All Reviewers

We thank all reviewers once again for their constructive feedback throughout the discussion. Based on the latest comments, we have uploaded a further revised version, with all changes marked in red. The main updates are summarized below.

  1. Framing as a benchmark paper. We present Real-ZAC and the task as the main contribution, and Pinpoint consistently as a baseline. Wording that emphasized the superiority of Pinpoint has been removed from the Introduction, the contribution list, the result discussions, and the Conclusion.
  2. Reorganization of the main text. The detailed formulation of Pinpoint has been moved to Appendix B, and the main text keeps a concise description of the baseline. Section 3 now includes the defect design criteria, the annotation and quality-control protocol, and representative defect examples (Figure 3).
  3. Task definition. The definition of Zero-shot Anomaly-Aware Counting is now stated more explicitly.
  4. Defect-state performance. Table 2 now reports the F1 on defective queries alone next to the pooled F1 for all compared methods.

We hope these revisions address the remaining concerns, and we sincerely appreciate the reviewers' time and effort.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.