Beyond Dents and Scratches: Logical Constraints in Unsupervised Anomaly Detection and Localization
Open accessAbstract
The unsupervised detection and localization of anomalies in natural images is an intriguing and challenging problem. Anomalies manifest themselves in very different ways and an ideal benchmark dataset for this task should contain representative examples for all of them. We find that existing datasets are biased towards local structural anomalies such as scratches, dents, or contaminations. In particular, they lack anomalies in the form of violations of logical constraints, e.g., permissible objects occurring in invalid locations. We contribute a new dataset based on industrial inspection scenarios that evenly covers both types of anomalies. We provide pixel-precise ground truth data for each anomalous region and define a generalized evaluation metric that addresses localization ambiguities that can arise for logical anomalies. Furthermore, we propose a novel algorithm that improves over the state of the art in the joint detection of structural and logical anomalies. It consists of a local and a global network branch. The first one inspects confined regions independent of their spatial locations in the input image and is primarily responsible for the detection of entirely new local structures. The second one learns a globally consistent representation of the training data through a bottleneck that enables the detection of violations of long-range dependencies, a key characteristic of many logical anomalies. We perform extensive evaluations on our new dataset to corroborate our claims.
AI snapshot
Written from the paper’s full text.
TL;DR
The paper introduces MVTec LOCO AD, a new industrial-inspection benchmark that balances structural anomalies (e.g., dents, scratches) with logical anomalies (e.g., missing, extra, or misplaced objects), plus the sPRO metric for evaluating localization under logical-anomaly ambiguity. It also proposes GCAD, a two-branch method that improves joint structural/logical anomaly detection and localization over prior methods.
Key findings
- MVTec LOCO AD contains 3644 images across five object categories, with 1772 training, 304 validation, and 1568 test images; it includes 89 defect types and pixel-precise ground truth.
- On MVTec LOCO AD, GCAD achieved mean sPRO 0.701, outperforming Student-Teacher (0.626), SPADE (0.451), VAE (0.382), AE (0.378), MNAD (0.339), f-AnoGAN (0.334), and Variation Model (0.225).
- GCAD was more balanced across anomaly types: structural sPRO 0.692 and logical sPRO 0.711, whereas Student-Teacher scored 0.756 structural but only 0.497 logical.
- On MVTec AD, GCAD reached mean sPRO 0.789 (structural 0.716, logical 0.863), compared with SPADE 0.640 and Student-Teacher 0.590; image-level AU-ROC on MVTec AD was 0.931 for GCAD versus 0.902 for SPADE.
Methods
- The authors created MVTec LOCO AD from five industrial inspection categories (breakfast box, screw bag, pushpins, splicing connectors, juice bottle), providing anomaly-free training/validation images and test images with structural and logical defects.
- They introduced saturated per-region overlap (sPRO), a generalization of PRO that saturates once a prediction overlaps a ground-truth region beyond a per-defect saturation threshold; they report area under the sPRO curve up to a false positive rate of 0.05.
- GCAD combines a local branch, based on a pretrained Student-Teacher-style local feature encoder and a regression network, with a global branch that learns a globally consistent representation through a low-dimensional bottleneck (g = 32) and a high-capacity regression network; anomaly maps from both branches are normalized and combined, with multi-scale evaluation using receptive fields p ∈ {17, 33}.
Limitations
- The paper states that GCAD can fail on very small anomalies, on challenging logical constraints such as enforcing a fixed number of objects that may appear almost anywhere, and on subtle/intricate differences from anomaly-free images.
- The paper states that saturation thresholds for sPRO were manually selected for each defect type; sensitivity analysis varied them from 0.5 to 1.5 times the original threshold and found rankings mostly stable.
- The paper states that sPRO does not account for false positive predictions, so it is reported together with false positive rate; the authors also note that due to the complexity of the new dataset, there remains room for future improvement.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 13 cited sources checked · 0 supported · 0 flagged
- Findingown data
The new dataset contains 3644 images of five distinct object categories inspired by real-world industrial inspection scenarios.
“It contains 3644 images of five distinct object categories inspired by realworld industrial inspection scenarios.”
Quote in the paperSubstantial claim - Findingnone given
In the new dataset, structural anomalies occur as scratches, dents, or contaminations in the manufactured products.
“Structural anomalies occur as scratches, dents, or contaminations in the manufactured products.”
Quote in the paperNot a substantial claim - Findingown data
Existing methods perform very differently on structural and logical anomaly classes.
“we classify an anomaly as either a structural anomaly or a logical anomaly and demonstrate that existing methods indeed perform very differently on these two classes.”
Quote in the paperSubstantial claim - Findingreasoning
The proposed algorithm's global network branch learns a globally consistent representation of the training data through a bottleneck that enables the detection of violations of long-range dependencies, a key characteristic of many logical anomalies.
“The second one learns a globally consistent representation of the training data through a bottleneck that enables the detection of violations of long-range dependencies, a key characteristic of many logical anomalies.”
Quote in the paperSubstantial claim - Findingown data
The paper proposes a novel algorithm that improves over the state of the art in the joint detection of structural and logical anomalies.
“we propose a novel algorithm that improves over the state of the art in the joint detection of structural and logical anomalies.”
- Also stated in: § 1 Introduction · plain
“We propose a new method for the unsupervised pixelprecise localization of anomalies.It improves the results of the joint detection of structural and logical anomalies compared to existing methods.”
- Also stated in: § 4 Description of Our Method · plain
“we introduce GCAD (Global Context Anomaly Detection), a new method for the unsupervised localization of anomalies that improves the joint detection of structural and logical anomalies compared to existing methods.”
Quote in the paperSubstantial claim - Also stated in: § 1 Introduction · plain
- Findingown data
The new dataset provides pixel-precise ground truth data for each anomalous region.
“We provide pixel-precise ground truth data for each anomalous region”
- Also stated in: § 3.2 Annotations and Labeling Policies · plain
“For all anomalies present in the dataset, we provide pixelprecise ground-truth annotations.”
Quote in the paperSubstantial claim - Also stated in: § 3.2 Annotations and Labeling Policies · plain
- Findingown data
The paper contributes a new dataset based on industrial inspection scenarios that evenly covers both structural and logical anomalies.
“We contribute a new dataset based on industrial inspection scenarios that evenly covers both types of anomalies.”
- Also stated in: § 2.1 Datasets · plain
“we introduce the Logical Constraints Anomaly Detection dataset.It represents industrial inspection scenarios and equally covers both types of anomalies.”
Quote in the paperSubstantial claim - Also stated in: § 2.1 Datasets · plain
- Findingown data
The SPADE method has difficulties localizing the more subtle logical anomaly in the cable image.
“The SPADE method manages to detect the missing transistor and performs well on the structural anomalies, but has difficulties to localize the more subtle logical anomaly in the image of the cable.”
Quote in the paperNot a substantial claim - Findingnone given
The Variation Model fails to localize the logical anomaly on the cable.
“However, it fails to localize the logical anomaly on the cable.”
Quote in the paperNot a substantial claim - Findingnone given
The Variation Model yields increased anomaly scores on the bent wires.
“It also yields increased anomaly scores on the bent wires.”
Quote in the paperNot a substantial claim - Findingown data
Our method might fail when anomalies are very small in size.
“Our method might fail when anomalies are very small in size, e.g., for the broken pushpin in the top left compartment.”
Quote in the paperNot a substantial claim - Findingown data
Combining multiple receptive fields together enhanced the performance for both structural and logical anomaly detection.
“When combining multiple receptive fields together, the performance for both structural and logical anomaly detection could be enhanced.”
Quote in the paperNot a substantial claim - Findingown data
Our approach performed equally well in the detection of structural and logical anomalies.
“Our approach performed equally well in the detection of structural and logical anomalies and improved the state of the art in the joint detection of both.”
Quote in the paperSubstantial claim - Findingown data
MNAD does not show a tendency to detect both of the logical anomalies.
“This is not the case for MNAD”
Quote in the paperNot a substantial claim - Findingown data
The results on the MVTec AD dataset are similar to the results on the MVTec LOCO AD dataset.
“The results are similar to the ones on the MVTec LOCO AD dataset.”
Quote in the paperNot a substantial claim- Rests on an assumption: The 37 MVTec AD anomalous test images identified by the authors contain defects that match their definition of logical anomalies.
- Findingown data
The Student-Teacher model performs well on the structural anomalies.
“The Student-Teacher model performs well on the structural anomalies but entirely fails to localize the logical anomalies due to its limited receptive field.”
Quote in the paperNot a substantial claim - Findingown data
The best balance between the detection of structural and logical anomalies is achieved for g = 32.
“While the mean performance is slightly better for g = 64, the best balance between the detection of structural and logical anomalies is achieved for g = 32.”
Quote in the paperNot a substantial claim - Findingown data
Combining A_loc and A_glo improved the performance for both structural and logical anomalies.
“Combining A loc and A glo improved the performance for both structural and logical anomalies.”
Quote in the paperSubstantial claim - Findingnone given
We developed a new method that permits the joint localization of both anomaly types.
“In addition, we developed a new method that permits the joint localization of both anomaly types.”
Quote in the paperSubstantial claim- Rests on an assumption: The first branch is based on an existing method that excels at the localization of structural anomalies.
- Findingown data
Our method might fail to capture very challenging logical constraints, such as enforcing a fixed number of objects that can potentially appear almost anywhere in the input image.
“It might also fail to capture very challenging logical constraints, such as enforcing a fixed number of objects that can potentially appear almost anywhere in the input image.”
Quote in the paperNot a substantial claim - Findingreasoning
Certain logical anomalies can be detected by the local branch, e.g., two pushpins being present in a single compartment, since both pushpins fall into the receptive field of E_loc.
“Certain logical anomalies can be detected by the local branch, e.g., two pushpins being present in a single compartment, since both pushpins fall into the receptive field of the local feature extractor E loc .”
Quote in the paperSubstantial claim - Findingown data
The ranking of the evaluated methods was stable across all ten runs with varied saturation thresholds, except for two runs in which two methods switched between the sixth and seventh rank.
“The ranking of the evaluated methods was stable across all ten runs, with the exception of two runs in which two methods switched between the sixth and seventh rank.”
Quote in the paperNot a substantial claim - Findingown data
Methods that leverage feature descriptors from pretrained networks, i.e., SPADE and Student-Teacher, outperform generative methods based on autoencoders or GANs.
“We also observe that methods that leverage feature descriptors from pretrained networks, i.e., SPADE and Student-Teacher, outperform the generative methods based on autoencoders or GANs.”
Quote in the paperNot a substantial claim - Findingreasoning
We created a new dataset for the unsupervised localization of anomalies that focuses on the detection of both structural and logical anomalies.
“We therefore created a new dataset for the unsupervised localization of anomalies that focuses on the detection of both structural and logical anomalies.”
Quote in the paperSubstantial claim- Rests on an assumption: Anomalies in natural images can manifest themselves in many different ways.
- Rests on an assumption: Previous work predominantly concentrated on the development of datasets and methods for the detection of structural anomalies.
- Findingreasoning
The Student-Teacher approach's performance dropped significantly for the logical anomalies in the dataset.
“However, its performance dropped significantly for the logical anomalies in the dataset.”
Quote in the paperSubstantial claim - Findingown data
Our proposed method performs slightly worse in the detection of structural anomalies than the Student-Teacher method and SPADE on the MVTec AD dataset.
“While our proposed method performs slightly worse in the detection of structural anomalies than the Student-Teacher method and SPADE, it excels in the classification of the logical anomalies on this dataset.”
Quote in the paperSubstantial claim - Findingown data
The SPADE method manages to detect the missing transistor.
“The SPADE method manages to detect the missing transistor and performs well on the structural anomalies, but has difficulties to localize the more subtle logical anomaly in the image of the cable.”
Quote in the paperNot a substantial claim - Findingnone given
All methods based on autoencoders tend to yield increased anomaly scores on the structural anomalies.
“All methods based on autoencoders tend to yield increased anomaly scores on the structural anomalies.”
Quote in the paperNot a substantial claim - Findingreasoning
Evaluating the reconstruction error of E_glo with respect to pretrained features after upsampling performed significantly worse than the feature regression approach for detecting logical anomalies.
“Figure 12 shows that evaluating the reconstruction error performed significantly worse than our feature regression approach.”
Quote in the paperSubstantial claim - Findingown data
Our method reliably detects all four defects.
“Our method reliably detects all four defects.”
Quote in the paperSubstantial claim - Findingown data
The SPADE method performs well on the structural anomalies.
“The SPADE method manages to detect the missing transistor and performs well on the structural anomalies, but has difficulties to localize the more subtle logical anomaly in the image of the cable.”
Quote in the paperNot a substantial claim - Findingreasoning
The Student-Teacher model entirely fails to localize the logical anomalies.
“The Student-Teacher model performs well on the structural anomalies but entirely fails to localize the logical anomalies due to its limited receptive field.”
Quote in the paperSubstantial claim - Findingnone given
For the missing transistor, f-AnoGAN manages to capture the logical constraint that a transistor should always be present.
“For the missing transistor, however, it manages to capture the logical constraint that a transistor should always be present.”
Quote in the paperNot a substantial claim - Findingown data
Progressively fading out the weights of the skip connections facilitated optimization, yielding lower values of L_kd on the training and validation set.
“We empirically observed that progressively fading out the weights of the skip connections facilitated the optimization, yielding lower values of L kd on the training and validation set.”
Quote in the paperNot a substantial claim - Findingown data
All other methods show a balanced classification performance between logical and structural anomalies at the image level.
“All other methods show a balanced classification performance between logical and structural anomalies.”
Quote in the paperNot a substantial claim - Findingreasoning
f-AnoGAN yields many false positives on areas that are challenging to reconstruct.
“Similarly to the autoencoders, f-AnoGAN yields many false positives on areas that are challenging to reconstruct.”
Quote in the paperNot a substantial claim - Findingown data
Our proposed method excels in the classification of the logical anomalies on the MVTec AD dataset.
“While our proposed method performs slightly worse in the detection of structural anomalies than the Student-Teacher method and SPADE, it excels in the classification of the logical anomalies on this dataset.”
Quote in the paperSubstantial claim - Findingnone given
Pixel-precise ground truth annotations are provided for each anomalous test image in the new dataset.
“Pixel-precise ground truth annotations are provided for each anomalous test image.”
Quote in the paperSubstantial claim - Findingown data
Existing methods tend to be biased towards the detection of one of the two types of anomalies.
“Our results showed that existing methods tend to be biased towards the detection of one of the two types of anomalies.”
Quote in the paperSubstantial claim - Findingown data
For the juice bottle, the Variation Model managed to detect parts of the structural anomaly.
“For the juice bottle, it managed to detect parts of the structural anomaly.”
Quote in the paperNot a substantial claim - Findingnone given
Our proposed method successfully localizes the anomaly in both images in Figure 2.
“Our proposed method successfully localizes the anomaly in both images”
Quote in the paperNot a substantial claim - Findingown data
Distillation of features from pretrained networks into E_glo greatly enhanced the anomaly localization performance for both structural and logical anomalies.
“The plot on the right-hand side of Fig. 14 shows that the distillation of features from pretrained networks into E glo greatly enhanced the anomaly localization performance for both structural and logical anomalies.”
Quote in the paperSubstantial claim - Findingown data
There exist cases in which the global branch contributes to a better detection of structural anomalies.
“There also exist cases in which the global branch contributes to a better detection of structural anomalies.”
Quote in the paperSubstantial claim - Findingown data
The high-dimensional and detailed feature maps of E_loc can only be approximately reproduced by E_glo due to its low-dimensional bottleneck.
“However, our ablation studies show that the high-dimensional and detailed feature maps of E loc can only be approximately reproduced by E glo due to its low-dimensional bottleneck.”
Quote in the paperSubstantial claim - Findingown data
Our method outperformed all other methods at the combined detection of structural and logical anomalies.
“Our method outperformed all other methods at the combined detection of structural and logical anomalies.”
Quote in the paperSubstantial claim - Findingown data
All methods based on autoencoders produce many false positives in areas that are difficult to accurately reconstruct, such as reflections on the wires of the cable.
“However, they also produce many false positives in areas that are difficult to accurately reconstruct, i.e., the reflections on the wires of the cable.”
Quote in the paperNot a substantial claim - Findingown data
f-AnoGAN yielded numerous false positive predictions in areas that are difficult to accurately reconstruct.
“Similar to the deterministic and variation autoencoders, f-AnoGAN yielded numerous false positive predictions in areas that are difficult to accurately reconstruct.”
Quote in the paperNot a substantial claim - Findingown data
The deterministic and variational autoencoders both yielded large residuals in parts of the images that are challenging to reconstruct.
“The deterministic and the variational autoencoder both yielded large residuals in the parts of the images that are challenging to reconstruct, e.g., on the cables between the two splicing connectors.”
Quote in the paperNot a substantial claim - Findingown data
All evaluated methods except ours show a bias towards the detection of one type of anomaly.
“All evaluated methods except ours show a bias towards the detection of one type of anomaly.”
Quote in the paperSubstantial claim - Findingown data
Both the VAE and the deterministic AE show a tendency to detect both of the logical anomalies.
“Both the VAE and the deterministic AE show a tendency to detect both of the logical anomalies.”
Quote in the paperNot a substantial claim - Findingreasoning
To achieve a perfect score, a method is not necessarily required to predict the whole ground truth area as anomalous.
“To achieve a perfect score, however, a method is not necessarily required to predict the whole ground truth area as anomalous.”
Quote in the paperSubstantial claim - Findingnone given
The Student-Teacher approach performed slightly better at the detection of structural anomalies.
“The Student-Teacher approach performed slightly better at the detection of structural anomalies.”
Quote in the paperNot a substantial claim - Findingown data
Our approach improved the state of the art in the joint detection of both structural and logical anomalies.
“Our approach performed equally well in the detection of structural and logical anomalies and improved the state of the art in the joint detection of both.”
Quote in the paperSubstantial claim - Findingown data
Our GCAD method fails to localize some anomalies.
“Qualitative examples for which our GCAD method fails to localize anomalies”
Quote in the paperNot a substantial claim - Premisecitation
Existing datasets lack anomalies in the form of violations of logical constraints, e.g., permissible objects occurring in invalid locations.
“In particular, they lack anomalies in the form of violations of logical constraints, e.g., permissible objects occurring in invalid locations.”
- Also stated in: § 3 The Logical Constraints Anomaly Detection Dataset · plain
“None of them, however, set an explicit focus on the joint detection of structural and logical anomalies.”
Quote in the paperSubstantial claim0 of 3 sources checked - Also stated in: § 3 The Logical Constraints Anomaly Detection Dataset · plain
- Premisereasoning
The development of methods capable of detecting logical anomalies is hindered by the availability of suitable data.
“The development of methods that are capable of detecting logical anomalies is hindered by the availability of suitable data.”
Quote in the paperSubstantial claim - Premisereasoning
To date, there exists no comprehensive dataset that explicitly focuses on the detection of structural as well as logical anomalies and that requires a model to understand the underlying logical or geometrical relationships in the anomaly-free data.
“To date, there exists no comprehensive dataset that explicitly focuses on the detection of structural as well as logical anomalies and that requires a model to understand the underlying logical or geometrical relationships in the anomaly-free data.”
Quote in the paperSubstantial claim - Premisenone given
Commonly used metrics are not directly applicable to assess the capability of methods to detect logical anomalies.
“We find that commonly used metrics are not directly applicable to assess the capability of methods to detect logical anomalies.”
Quote in the paperSubstantial claim - Premisenone given
Logical anomalies occur in manufacturing processes, e.g., as an incorrect wiring of a circuit, a shift in the fill level of a vial, or the absence of an essential component.
“Logical anomalies, however, do occur in manufacturing processes, e.g., as an incorrect wiring of a circuit, a shift in the fill level of a vial, or the absence of an essential component.”
Quote in the paperNot a substantial claim - Premisecitation
Existing datasets for unsupervised anomaly detection and localization are biased towards local structural anomalies such as scratches, dents, or contaminations.
“We find that existing datasets are biased towards local structural anomalies such as scratches, dents, or contaminations.”
- Also stated in: § 2.1 Datasets · plain
“all of them are primarily designed for the detection of what we refer to as structural anomalies.”
- Also stated in: § 1 Introduction · plain
“all of them focus on the detection of structural anomalies and therefore favor methods that perform well on this type of anomaly.”
Quote in the paperSubstantial claim0 of 3 sources checked - Also stated in: § 2.1 Datasets · plain
- Premisecitation
The Student-Teacher framework's networks exhibit a limited receptive field, which prevents the method from detecting global inconsistencies that fall outside the receptive field's range.
“The networks employed exhibit a limited receptive field, which prevents this method from detecting global inconsistencies that fall outside the receptive field's range.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Bergmann et al. 2020
Not checked in the free run.
- Interpretationreasoning
The local branch of the model is expected to perform well in the detection of structural anomalies.
“Hence, we expect the local branch of our model to perform well in the detection of structural anomalies.”
Quote in the paperSubstantial claim - Interpretationreasoning
MNAD's high-capacity memory module allows it to reconstruct the areas that contain logical anomalies.
“for which the high-capacity memory module allows to reconstruct the areas that contain logical anomalies.”
Quote in the paperSubstantial claim - Interpretationreasoning
The difference in architecture between E_glo and R_glo is crucial to the method.
“The difference in architecture between E glo and R glo is crucial to our method.”
Quote in the paperSubstantial claim - Interpretationreasoning
The skip connections enable R_glo to solve the regression task without capturing the global context of the training data.
“The skip connections enable R glo to solve the regression task without capturing the global context of the training data.”
Quote in the paperSubstantial claim - Interpretationnone given
Large regression errors indicate anomalous pixels.
“Large regression errors indicate anomalous pixels.”
Quote in the paperSubstantial claim - Interpretationnone given
The choice of the receptive field p for E_loc can have a significant impact on anomaly localization performance, especially when anomalies vary greatly in size.
“The choice of the receptive field p for E loc can have a significant impact on the anomaly localization performance, especially when anomalies vary greatly in size.”
Quote in the paperNot a substantial claim - Interpretationnone given
The Student-Teacher model's failure to localize logical anomalies is due to its limited receptive field.
“The Student-Teacher model performs well on the structural anomalies but entirely fails to localize the logical anomalies due to its limited receptive field.”
Quote in the paperSubstantial claim - Interpretationreasoning
Choosing a latent dimension that is too large allows the global feature encoder to copy parts of its input directly into the latent representation.
“This is due to the fact that choosing a latent dimension that is too large allows the global feature encoder to copy parts of its input directly into the latent representation.”
Quote in the paperSubstantial claim - Interpretationreasoning
Since the weights of both E_glo and R_glo are randomly initialized, there exists no training incentive for the two networks to behave differently for structural anomalies.
“Since the weights of both E glo and R glo are randomly initialized, there exists no training incentive for the two networks to behave differently for structural anomalies.”
Quote in the paperNot a substantial claim - Interpretationreasoning
A direct comparison of E_loc features to U•E_glo features would lead to many false positives in anomaly images due to inaccurate feature reconstructions.
“A direct comparison would lead to many false positives in the anomaly images due to inaccurate feature reconstructions.”
Quote in the paperSubstantial claim - Interpretationreasoning
The local branch is inherently ill-suited for the detection of anomalies that violate long-range dependencies, which is characteristic of many logical anomalies such as missing or additional objects.
“Hence, our local branch is inherently ill-suited for the detection of anomalies that violate long-range dependencies, which is characteristic for many logical anomalies such as missing or additional objects in the input image.”
Quote in the paperSubstantial claim- Rests on an assumption: Many logical anomalies are characterized by violations of long-range dependencies, such as missing or additional objects.
- May combine two claims. 34
- Classified as probably not a substantial claim. 15
- The quote only partly states this claim. 1
May combine two claims.
Findingown dataIn the authors' experiments, MNAD helps in the detection of structural anomalies but impairs the detection of logical ones.
“In our experiments, we observed that this indeed helps in the detection of structural anomalies but impairs the detection of logical ones (see Fig. 8).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Park et al. 2020
Not checked in the free run.
May combine two claims.
FindingreasoningIn the toy dataset, a VAE identifies the two circles as anomalous but produces many false positives in the anomaly-free test image.
“This method accomplishes to identify the two circles as anomalous but produces many false positives in the anomaly-free test image.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenMVTec LOCO AD consists of five object categories from industrial inspection scenarios.
“MVTec LOCO AD consists of five object categories from industrial inspection scenarios.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Findingnone givenIn the new dataset, logical anomalies violate underlying constraints, e.g., a permissible object being present in an invalid location or a required object not being present at all.
“Logical anomalies violate underlying constraints, e.g., a permissible object being present in an invalid location or a required object not being present at all.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningThe proposed algorithm's local network branch inspects confined regions independent of their spatial locations and is primarily responsible for the detection of entirely new local structures.
“The first one inspects confined regions independent of their spatial locations in the input image and is primarily responsible for the detection of entirely new local structures.”
- Also stated in: § 4 Description of Our Method · plain
“Since this method computes anomaly scores for locally confined image regions independent of their spatial position in the input image, we refer to this branch as the local branch of our model.”
Quote in the paperSubstantial claim- Also stated in: § 4 Description of Our Method · plain
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Findingnone givenThe proposed performance metric is a generalization of an established measure for unsupervised anomaly detection.
“This performance measure is a generalization of an established measure for unsupervised anomaly detection.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataExtensive evaluations against state-of-the-art methods show the superiority of the proposed approach in the detection of logical anomalies, as well as in the combined localization of both anomaly types.
“Extensive evaluations against state-of-the-art methods show the superiority of our approach in the detection of logical anomalies, as well as in the combined localization of both anomaly types.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningIn the toy dataset, the Student-Teacher method clearly identifies and localizes the color defect but does not predict the two circles as anomalous.
“This method clearly identifies and localizes the color defect.The two circles, however, are not predicted as anomalous because each individual circle does not constitute an anomaly and the receptive field of the method is not large enough to understand the long-range relationships in the image.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Bergmann et al. 2020
Not checked in the free run.
Classified as probably not a substantial claim.
FindingreasoningThe paper defines a generalized evaluation metric that addresses localization ambiguities that can arise for logical anomalies.
“define a generalized evaluation metric that addresses localization ambiguities that can arise for logical anomalies.”
- Also stated in: § 1 Introduction · plain
“we introduce a performance metric that takes the different modalities of the defects present in our dataset into account.”
Quote in the paperSubstantial claim- Also stated in: § 1 Introduction · plain
May combine two claims.
FindingcitationThe MVTec AD dataset's vast majority (97%) of anomalies match the definition of structural anomalies, so evaluation on this dataset alone does not give sufficient insight into how well a method detects logical anomalies.
“The vast majority (97%) of anomalies in the dataset matches our definition of structural anomalies.Hence, an evaluation on this dataset alone does not give sufficient insight into how well a method detects logical anomalies.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites Bergmann et al. 2021 source
Not checked in the free run.
May combine two claims.
Findingown dataOur GCAD method performs significantly better than all other evaluated methods in the detection of logical and the joint detection of structural and logical anomalies at the image level.
“Similar to our experiments on anomaly localization, our GCAD method performs significantly better than all other evaluated methods in the detection of logical and the joint detection of structural and logical anomalies.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataMNAD reduced the number of false positive predictions and improved upon the basic autoencoders in the detection of structural anomalies, but it failed to detect logical anomalies.
“While the memory module in MNAD reduced the number of false positive predictions and improved upon the basic autoencoders in the detection of structural anomalies, it failed to detect the logical anomalies.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenNot all of the image augmentations are suited for every type of object in the dataset.
“Not all of these augmentations are suited for every type of object in our dataset.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataOur method yielded a similar mean performance for receptive fields of size 17 and 33, while the performance dropped for very large values of p.
“Our method yielded a similar mean performance for receptive fields of size 17 and 33, while the performance dropped for very large values of p.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenThe authors selected suitable saturation thresholds for each of the 89 individual defect types in the dataset.
“We selected suitable saturation thresholds for each of the 89 individual defect types that occur in our dataset.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataOur method performed well for various values of d_glo and is not highly sensitive to this parameter.
“The plot on the left-hand side of Fig. 14 indicates that our method performed well for various values of d glo and is not highly sensitive to this parameter.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThe Student-Teacher method performs best in the detection of structural anomalies at the image level, but its performance on logical anomalies is significantly lower.
“The Student-Teacher method performs best in the detection of structural anomalies, however, its performance on the logical ones is significantly lower.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThe training set consists of only anomaly-free images, and the validation images are free of any anomalies.
“The training sets consists of only anomaly-free images.”
- Also stated in: § 3.1 Description of the Dataset · plain
“Like the training images, the validation images are free of any anomalies.”
Quote in the paperSubstantial claim- Also stated in: § 3.1 Description of the Dataset · plain
Classified as probably not a substantial claim.
Findingnone givenWe introduced a new performance metric that takes the different modalities of the two anomaly types into account.
“Furthermore, we introduced a new performance metric that takes the different modalities of the two anomaly types into account.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataThe dataset provides a total of 1772 images for training, 304 for validation, and 1568 for testing.
“We provide a total of 1772 images for training, 304 for validation, and 1568 for testing.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingnone givenThe Variation Model manages to detect parts of the damaged transistor as well as its absence.
“The Variation Model manages to detect parts of the damaged transistor as well as its absence.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingreasoningThe global branch also detects the two-pushpins anomaly but tends to produce more false positive predictions than the local branch since it has to reconstruct the entire input image over a low dimensional bottleneck.
“The global branch also detects this anomaly.However, it also tends to produce more false positive predictions than the local branch since it has to reconstruct the entire input image over a low dimensional bottleneck.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataThe local branch is mainly responsible for detecting structural anomalies, while the global branch is mainly responsible for detecting logical anomalies.
“A loc is mainly responsible for detecting structural anomalies, while A glo enables the network to detect logical anomalies, as illustrated in Fig. 6.”
- Also stated in: § Scoring Functions for Anomaly Localization. · boosted
“Our experiments show that the global branch is indeed mainly responsible for the detection of logical anomalies, while the local branch performs much better at the detection of structural anomalies.”
- Also stated in: § Global Context Dimension. · plain
“We compared the performance of the local anomaly maps A loc to that of the global anomaly maps A glo and saw that, indeed, A loc performed much better in the detection of structural anomalies.”
Quote in the paperSubstantial claim- Also stated in: § Scoring Functions for Anomaly Localization. · boosted
May combine two claims.
FindingreasoningThe paper's labeling policy and evaluation metric take into account ambiguities in labeling logical defects.
“Our labeling policy and the newly introduced evaluation metric take such ambiguities into account.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Findingnone givenOur method performed well for all of the displayed examples.
“Our method performed well for all of the displayed examples.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThe Student-Teacher approach detected structural anomalies reliably but failed to detect logical anomalies due to its limited receptive field.
“While the Student-Teacher approach detected the structural anomalies reliably, it failed to detect the logical anomalies due to its limited receptive field.”
Quote in the paperSubstantial claimMay combine two claims.
Findingnone givenThe SPADE method failed to detect the flipped splicing connector, while it managed to localize the remaining three anomalies.
“The SPADE method, on the other hand, failed to detect the flipped splicing connector, while it managed to localize the remaining three anomalies.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThe local branch works better in the detection of a logical anomaly than the global branch in some cases, and vice versa.
“Qualitative examples for which the local branch works better in the detection of a logical anomaly than the global branch and vice versa.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Findingnone givenThe proposed method consists of two main branches, one primarily responsible for the localization of structural anomalies and the other for the localization of logical anomalies.
“Our method consists of two main branches, one of which is primarily responsible for the localization of structural anomalies and the other one for the localization of logical anomalies.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataOur method significantly outperforms all other approaches in the detection of logical anomalies, while maintaining a high performance at the detection of structural anomalies.
“Our method significantly outperforms all other approaches in the detection of logical anomalies, while maintaining a high performance at the detection of structural anomalies.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningWhen the global context dimension was increased, overall detection performance was not affected substantially, but there was a slight decline in the detection of logical anomalies while the localization of structural anomalies improved.
“When the global context dimension was increased, the overall detection performance was not affected substantially.However, we observed a slight decline in the detection of logical anomalies, while the localization of structural anomalies improved.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Findingnone givenEach dataset category possesses certain logical constraints, and violations to any of these constraints constitute logical anomalies.
“Each dataset category possesses certain logical constraints.”
- Also stated in: § 3.1 Description of the Dataset · plain
“Violations to any of these constraints constitute logical anomalies.”
Quote in the paperSubstantial claim- Also stated in: § 3.1 Description of the Dataset · plain
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingown dataWe observe seemingly better performances for all methods when increasing the integration threshold.
“We observe seemingly better performances for all methods when increasing the integration threshold.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThe test set contains anomaly-free images and images with various types of logical and structural anomalies.
“The test set contains anomaly-free images and images with various types of logical and structural anomalies.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningThe increased capacity of E_glo led to fewer false positives in the global anomaly detection branch while E_glo captured the global context of the data less reliably.
“The increased capacity of E glo led to fewer false positives in the global anomaly detection branch while, at the same time, E glo captured the global context of the data less reliably.”
Quote in the paperSubstantial claimThe quote only partly states this claim.
Findingown dataOur method consistently outperforms all other evaluated methods for all but one of the dataset categories on MVTec LOCO AD.
“Our method consistently outperforms all other evaluated methods for all but one of the dataset categories.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningThe Variation Model requires a pixel-precise alignment of the inspected objects; since this is not possible for the splicing connectors, it did not perform well for this dataset category.
“The Variation Model requires a pixel-precise alignment of the inspected objects.Since this is not possible for the splicing connectors, it did not perform well for this dataset category.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Findingnone givenIn the dataset, the union of all areas of the image that could potentially be the cause for the anomaly is labeled as anomalous.
“In our dataset, the union of all areas of the image that could potentially be the cause for the anomaly is labeled as anomalous.”
Quote in the paperSubstantial claimMay combine two claims.
PremisereasoningAutoencoder-based methods tend to produce blurry and inaccurate reconstructions, leading to an increase in false positives.
“autoencoders tend to produce blurry and inaccurate reconstructions.This leads to an increase in false positives.”
Quote in the paperSubstantial claimMay combine two claims.
PremisereasoningExisting state-of-the-art unsupervised anomaly detection methods that model local features are inherently limited to information inside the receptive field of their descriptors, making it difficult to detect anomalies that violate long-range dependencies.
“They are, however, inherently limited to the information inside the receptive field of their descriptors.This makes it difficult to detect anomalies that violate long-range dependencies.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites Bergmann et al. 2020
Not checked in the free run.
- cites Burlina et al. 2019
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationMethods that leverage features of pretrained networks tend to outperform autoencoder- or GAN-based methods that are trained from scratch.
“Methods that leverage features of pretrained networks tend to outperform autoencoder-or GAN-based methods that are trained from scratch (Burlina et al. 2019).”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites Burlina et al. 2019
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone givenThe phenomenon of a too-large latent dimension allowing copying of input into the latent representation can also be observed in other bottleneck architectures, such as autoencoders.
“This phenomenon can also be observed in other bottleneck architectures, such as autoencoders.”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductown dataThe training, validation, and test sets are independent in the sense that they consist of images of distinct physical objects and there is no overlap between them.
“All three sets are independent of each other in the sense that they consist of images of distinct physical objects and that there is no overlap between them.”
Quote in the paperSubstantial claimMay combine two claims.
InterpretationreasoningThe high capacity of R_glo allows it to accurately reproduce the features of E_glo, which reduces the number of false positive detections compared to the reconstruction error between E_glo and U•E_glo.
“The high capacity of R glo allows it to accurately reproduce the features of E glo , which reduces the number of false positive detections compared to the reconstruction error between E glo and U • E glo .”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationnone givenMethods that compress the input data to a low-dimensional bottleneck possess the ability to capture logical constraints and fail to reproduce input images that violate them.
“Its design is inspired by the observation in Fig. 1 that methods that compress the input data to a low-dimensional bottleneck possess the ability to capture logical constraints and fail to reproduce input images that violate them.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationown dataSome logical anomalies are better detected by the local branch and some structural anomalies by the global branch.
“This indicates that some of the logical anomalies are better detected by the local branch and some structural ones by the global branch.”
Quote in the paperSubstantial claimMay combine two claims.
InterpretationreasoningThe reconstruction of 128-dimensional pretrained features through a small bottleneck is challenging and leads to many false positives.
“This is because the reconstruction of 128-dimensional pretrained features through a small bottleneck is challenging and leads to many false positives.”
Quote in the paperSubstantial claim
The paper relies on this without support.
Many logical anomalies are characterized by violations of long-range dependencies, such as missing or additional objects.
“which is characteristic for many logical anomalies such as missing or additional objects in the input image.”
The paper relies on this without support.
The validation set of the dataset contains only anomaly-free images.
“Note that the validation set of our dataset only contains anomaly-free images.”
The paper relies on this without support.
The 37 MVTec AD anomalous test images identified by the authors contain defects that match their definition of logical anomalies.
“Of the 1258 anomalous test images, we identified 37 to contain defects that match our definition of logical anomalies.”
The paper relies on this without support.
Anomalies in natural images can manifest themselves in many different ways.
“This paper is based on the observation that anomalies in natural images can manifest themselves in many different ways.”
The paper relies on this without support.
Previous work predominantly concentrated on the development of datasets and methods for the detection of structural anomalies.
“Previous work predominantly concentrated on the development of datasets and methods for the detection of structural ones.”
The paper relies on this without support.
The first branch is based on an existing method that excels at the localization of structural anomalies.
“The first is based on an existing method that excels at the localization of structural anomalies.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Since the release of the MVTec 3D-AD dataset, several papers have focused on anomaly detection in 3D industrial images.
“Since the release of MVTec 3D-AD [6] dataset, several papers have focused on anomaly detection in 3D industrial images.”
From Deep Industrial Image Anomaly Detection: A Survey (2024) · cites this paper as [6]
Section 3.1 (page unavailable) identifies the dataset introduced by this paper as MVTec LOCO AD, not MVTec 3D-AD. The cited paper therefore does not support the sentence’s identification of it as the MVTec 3D-AD dataset release; the passages also do not address the claimed subsequent trend in 3D-image research. (LLM, conf=0.87)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Global-branch reconstruction error names the wrong features
MinorThe text describes a reconstruction error between E glo and U • E glo, but the defined reconstruction loss compares E loc with U(E glo). The former pair also has different feature dimensions under the stated architecture.
“The high capacity of R glo allows it to accurately reproduce the features of E glo , which reduces the number of false positive detections compared to the reconstruction error between E glo and U • E glo .”
“L kd (I ) = E loc (I ) -U (E glo (I )) 2 F .”
Figure 10 assigns the third row to both anomaly types
MinorThe caption identifies the third row as a structural-anomaly example and also as a logical-anomaly example, although it describes the four rows as separate examples of the two types.
“The first and third row contain examples of structural anomalies, i.e. the flipped connector and the contamination in the juice bottle.”
“The second and third row contain examples of logical anomalies, i.e., a second cable being present between the two connectors and the banana label on the bottle filled with orange juice”
The citation lists haven’t been fetched yet
References 31 · Cited by 286
Research works printed but not linked: 4
- An, J., & Cho, S. (2015). Variational autoencoder based anomaly detection using reconstruction probability. Special Lecture on IE, 2(1), 1–18.
- Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative Adversarial Nets. In Advances in Neural Information Processing Systems, pp 2672–2680.
- Kingma, D.P., Ba, J. (2015). Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations (ICLR).
- Steger, C., Ulrich, M., & Wiedemann, C. (2018). Machine Vision Algorithms and Applications (2nd ed.). Wiley-VCH.
- Authors
- Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., & Steger, C.
- Affiliations
- Technical University of Munich (DE)
- Corresponding author
- Bergmann, Paul
- Venue
- International Journal of Computer Vision
- Published
- 2022-02-22
- Volume
- 130
- Issue
- 4
- Pages
- 947-969
- ISSN
- 0920-5691 (ISSN-L)
- 1573-1405
- OpenAlex
- W4212874935
- Cited by
- 286
- References
- 31
- Citations by year
- 2026: 110 · 2025: 87 · 2024: 69 · 2023: 18 · 2022: 2
- FWCI
- 20.19
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Locations
- International Journal of Computer Vision (publishedVersion, cc-by)
- mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) (submittedVersion, other-oa)
- Topics
- Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
- Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
- Keywords
- unsupervised anomaly detection, anomaly localization, logical constraints, structural anomalies, industrial inspection, evaluation metrics, long-range dependencies, natural images
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
