CostFilter-AD: Enhancing Anomaly Detection through Matching Cost Filtering

Q1: The proposed method is simple, effective, and well-proven. Calculation cost and memory cost can be added.

A1: Thank you for acknowledging our work's effectiveness and suggesting the inclusion of computational and memory costs. We include comparisons of parameter count, FLOPs, memory usage, inference time, and overall training time in Table R6, showing that CostFilter-AD introduces marginal overhead while consistently improving AD performance (Table R7).

Table R6. Comparison of computational and memory costs. ''-'' denotes training-free.

Table R7. Experimental comparison with baselines and +Ours.

Q2: Concern over the performance gaps compared to SOTA methods and some performance decrease.

A2: Thank you for raising this concern.

The performance gains of our method is crystal clear. To address the confusion, we would like to provide the following clarifications:

1. Fair experimental setup: All comparisons (baseline/+Ours) are conducted with identical image resolutions and template counts to ensure fair evaluation.

2. Comprehensive validation: As shown in Table R7, CostFilter-AD was applied to a range of recent multi-class anomaly detection methods, including UniAD (NeurIPS’22), GLAD (ECCV’24), HVQ-Trans (NeurIPS’23), AnomalDF (WACV’25), and Dinomaly (CVPR’25). We evaluated these models on standard benchmarks such as MVTec-AD, VisA, MPDD, and BTAD. Consistent improvements were observed in category-averaged metrics, see the link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf) for more details.

3. SOTA performance: When integrated with Dinomaly, CostFilter-AD achieves the best performance across all evaluation metrics on the four benchmarks. For example, AUROC scores (image/pixel) reach 99.7%/98.4% on MVTec-AD, 98.7%/98.8% on VisA, 97.5%/99.2% on MPDD, and 95.5%/98.1% on BTAD.

4. Clarifying the performance gap: The gap you mentioned arise from different settings in image resolution and number of templates, particularly between AnomalDF (+Ours) and AnomalyDino, also noted by Reviewer kyJ7. This discrepancy is due to our method using a lightweight configuration (224×224 resolution, 3 template images per test sample), whereas AnomalDF uses higher resolutions (448 or 672) and utilizes all training images as templates. Further details are in responses A2 and A3 to Reviewer kyJ7.

5. Category performance VS average performance: While there may be minor fluctuations in performance within certain categories, our multi-class AD model substantially improves average metrics across datasets.

6. Strong qualitative support: In our submission, we provide comprehensive qualitative results in Figures 1, 3, 5, 8, 9, 10, and 11, clearly demonstrating the effectiveness and adaptability of our method.

We sincerely hope that our clarifications can address your concerns. If have any remaining questions, please let us know, and we will do our best to address them.

Table R5. AnomalDF (abbr. as ADF) /+Ours Comparison under a fair setting.

Q1: Discrepancy between AnomalDF (AnomalyDINO) performance and reported results.

A1: Thank you! The gap is primarily due to two factors:

• Number of templates: In our experiments, we re-run AnomalyDINO using 3 randomly sampled images from the training dataset as reference templates during testing. In contrast, the original full-shot setting of AnomalyDINO employed the entire training dataset as templates.
• Image resolution: We resize images to 256×256, whereas the original AnomalyDINO resized to 448×448. Higher resolution allows the model to capture more details.

As shown in the Table R5, under AnomalyDINO's full-shot setting, we achieve image AUROC scores of 99.5 (MVTec-AD) and 97.4 (VisA), outperforming AnomalDF's 99.3/97.2.

Q2: Clarification on the 3-shot setting: Is the AnomalDF+Ours actually few-shot?

A2: The answer is No. The AnomalDF (+Ours) in our paper is full-shot, but differs from the original AnomalyDINO setting.

• Training: AnomalDF (+Ours) uses a fixed number (N=3) of templates per input image, randomly sampled from the full training set for each input, rather than drawn from a fixed template set as in the original few-shot setting of AnomalyDINO. Our dynamic sampling ensures the template pool covers the full training distribution; thus, we classify it as full-shot, offering a trade-off between template diversity and memory efficiency.
• Test: For fairness, we evaluate AnomalDF (+Ours) using our dynamic 3-shot sampling protocol, as reflected in the results reported in our submission.

Q3: Storage overhead claim.

A3: Thanks for your kindly reminding. We will carefully revise our statement on “storage overhead” to provide a more accurate and precise explanation.

Q4: Evaluation of full-shot AnomalyDINO.

A4: Following your advice, we tested under the original full-shot AnomalyDINO setting. In Table R5, Exp. 3–8 report results on MVTec-AD and Exp. 10–16 on VisA. CostFilter-AD consistently improves AnomalDF’s performance across all resolutions, with some cases (e.g., Exp. 6 vs. 7) showing that AnomalDF+Ours at lower resolutions matches or surpasses the baseline at higher resolutions.

We are integrating CostFilter-AD with PatchCore, similar to AnomalDF (+Ours), and will report it in the revised version.

Q5: How to utilize the reconstructed features from HVQ-Trans.

A5: Thanks! We directly use the input and reconstructed features from HVQ-Trans to construct the cost volume, without decoding them back to the image domain, as the HVQ-Trans already provides the necessary feature representations. This differs from Glad+Ours, which uses a pre-trained encoder to extract image features. We will state the implementation of HVQ-Trans+Ours more clearly in the revised vision.

Q6: Following the above, can CostFilter-AD be ensembled on feature-reconstruction methods like RD4AD, UniAD, MambaAD, Dinomaly?

A6: Yes! Similar to HVQ-Trans (+Ours), for feature-reconstruction-based methods, we bypass the feature encoder and directly construct the cost volume using both input and reconstructed features. We further validated it by integrating with UniAD and Dinomaly on MVTec-AD and VisA. SOTA results show CostFilter-AD's generalizability (see Table R7 below and the link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf) for details).

Q7: More datasets.

A7: Thanks! We have extended our evaluation by integrating CostFilter-AD with HVQ-Trans and Dinomaly on two more datasets: BTAD and MPDD. As shown in Table R7 and the link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf), we consistently improve baseline performance, validating effectiveness.

We sincerely appreciate your feedback and hope our response clarifies your concerns. If you have further questions, please feel free to let us know, and we'll be glad to clarify.

Q1: The paper is supported by solid claims and motivations. It could be better if the authors could apply the proposed method plug-in design into multiple AD baselines.

A1: Thanks! We apply our method to UniAD (NeurIPS'22) (new add), GLAD (ECCV'24), HVQ-Trans (NeurIPS'23), AnomalDF (WACV'25), Dinomaly [r1] (CVPR'25) (new add) on benchmarks MVTec-AD, VisA, BTAD (new add), and MPDD (new add). Please refer to Table R1 for category-averaged metrics, with detailed per-category metrics available in the link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf).

[r1] Jia Guo, et al. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. CVPR 2025.

Table R1. Evaluation with more baselines on multiple benchmarks via 7 comprehensive metrics.

Q2: It would be better if the authors also report metric such as PR, or AUPRC.

A2: Thank you! We would like to clarify that AUPRC (i.e., the area under the precision-recall curve) has already been reported in Tables 7–10 of our submission, where we refer to it as I-AP (image-level AUPRC) and P-AP (pixel-level AUPRC), following the terminology adopted by Glad, AnomalDF, and Dinomaly. The AUPRC is computed by the average_precision_score function from sklearn.metrics. We will clarify this in the revised version to avoid confusion.

Q3: Some of the SOTA methods have been missing in the paper.

A3: Thank you! We acknowledge the importance of the SOTA methods you mentioned and will incorporate both citations and discussions in the revised version for a more thorough and balanced evaluation.

Q4: The multiple steps may increase the difficulty of implementation. Could the authors provide how the model sensitive to architectures or hyper parameters selections.

A4: Thanks!

• To address potential concerns about implementation and reproducibility, we will release the full source code and model weights.

• We have also conducted comprehensive ablation studies in Table 4 to examine the model's sensitivity to architectural choices. Specifically, we evaluate different cost volume constructions (DN→depth/channel), template selection strategies ($C_0$ vs. $C_{N-1}$), dual-stream attention (SG and MG), and class-aware adaptation (loss $\mathcal{L}_s$). Results confirm that each component plays a vital role in achieving strong performance, validating their necessity and effectiveness.

Q5: Also it would be valuable for authors to compute the complexity like inference/training time.

A5: Thanks! We have provided the inference time and memory usage in Table 6 of the submission. These metrics, along with overall training time, are provided in Table R2 below. As shown, our method incurs reasonable computational overhead and minimal memory usage, while delivering notable performance improvements.

Table R2. Complexity comparison of baselines and +Ours. ''-'' denotes training-free.

We sincerely appreciate your feedback and hope our response has addressed your concerns. If you have any additional questions, please let us know, and we will respond promptly.

Q1: Missing the first multi-class anomaly detection work, UniAD, and a performance comparison.

A1: Thanks for your reminding.

• We fully recognize UniAD (NeurIPS'22) as a pioneering work in multi-class anomaly detection. In response, we have conducted extensive evaluations by integrating CostFilter-AD into UniAD, and will include both citation and performance comparisons in the revised paper.
• As shown in Table R3, on MVTec-AD, this integration improves image-level AUROC/AUPRC/F1-max by +1.5%/+0.6%/+1.1%, and pixel-level AUROC/AUPRC/F1-max/AUPRO by +0.6%/+16%/+9.4%/+0.7%, respectively. On VisA, we observe gains of +0.6%/+0.4%/+0.4% at the image level and +0.6%/+1.3%/+0.6%/+10.3% at the pixel level.
• Furthermore, we incorporated CostFilter-AD into Dinomaly (CVPR’25) and validated its effectiveness across more benchmarks. These consistent gains further highlight the effectiveness, flexibility, and generalizability of our method. Please refer to the link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf) for more category-aware details.

Table R3. Integration of our CostFilter-AD with UniAD and Dinomaly baselines.

Q2: Besides detection performance, parameter efficiency is crucial. MOE-AD highlights this in multi-class detection. The authors should provide results to demonstrate the efficiency and effectiveness.

A2: Thank you for highlighting the parameter efficiency.

• MoE-AD presents an elegant solution by significantly reducing model size via recursive ViT blocks and MoE-based FFN selection, setting a strong benchmark for balancing accuracy and efficiency in resource-constrained scenarios. We will cite this excellent work in our revised paper and are excited to explore its potential synergy with CostFilter-AD in future work.

• In response, we present a detailed comparison across multiple baselines and benchmarks (see Table R4), reporting parameters, FLOPs, memory usage, and inference time. In terms of memory usage, the increase is negligible. Regarding computational overhead, it remains relatively low compared to diffusion-based methods. For other approaches, the overhead can be further minimized by converting global matching into local matching, thereby optimizing the cost volume for more efficient computation. Regarding efficiency, our method provides notable performance gains with a reasonable increase in inference time.

Table R4. Parameter efficiency comparison and performance gain of baselines/+Ours on MVTec-AD.

Notably, the #Params in our model varies slightly across different baselines. This is because we need to map the matching features, which have different #Channels (e.g., 196, 768, or 1024), into a unified 96-dimensional space.

Q3: It would be beneficial if the authors could provide failure cases to enable a deeper analysis.

A3: Thank you for the insightful suggestion. As illustrated in Fig. 1 (at link (https://anonymous.4open.science/r/ICML-ID8276/PDF.pdf)), we present representative failure cases from six categories in MVTec-AD and VisA, demonstrating the performance of our method before and after filtering.

While our approach effectively reduces matching noise, its effectiveness depends on the presence of anomaly-relevant signals in the cost volume. If these signals are absent—due to low-resolution inputs or insufficient feature learning—the filtering process cannot recover them. In other words, as a denoising rather than a generative module, CostFilter-AD enhances existing features but cannot infer anomalies from missing evidence.

We appreciate your feedback and hope that our revisions address your concerns. If there are any remaining issues or questions, please let us know, and we will respond promptly.