Pushing the Limits of Fewshot Anomaly Detection in Industry Vision: Graphcore

• Changeover can degrade anomaly detection when production equipment is not fully fine-tuned, while large-data training slows adaptation to new products.Source: §1 · p. 1
• Industrial images have limited shape and color diversity but may vary in shooting angle or rotation, motivating rotation-invariant structural features.The introduction argues that CNN features lack rotation invariance, so multiple rotations of the same patch may create redundant memory-bank entries.Source: §1 · p. 2
• GraphCore is intended to train rapidly on a small number of normal samples for a new product, without transferring anomalies from an old product.Source: §1 · p. 2
• GraphCore uses visual isometric invariant features as anomaly measurements to reduce redundant features in memory-bank-based detection.The paper also presents feature augmentation to investigate CNN-generated features.Source: §1 · p. 2
• VIIFs significantly enhance few-shot anomaly detection performance on MVTec AD and MPDD.Source: §1 · p. 3
• Prior industrial few-shot approaches use abnormal test images, require many base-class images for meta-learning, or are not designed for changeover settings.The passage names Liznerski et al. (2020), Pang et al. (2021), and Ding et al. (2022) for abnormal-image testing; Wu et al. (2021) and Huang et al. (2022) for meta-learning; and PatchCore (Roth et al. (2022)), SPADE (Cohen & Hoshen (2020)), and PaDiM (Defard et al. (2021)) for few-shot MVTec AD settings not intended for changeover.Source: §1 · p. 3

2.1 AUGMENTATION+PATCHCORE

• Aug.(R) greatly outperforms SOTA models under the proposed few-shot setting.It augments data by rotation before extracting features from the ImageNet-pretrained model.Source: §2.1 · pp. 3–4

2.2 VISION ISOMETRIC INVARIANT FEATURE

• VIIG extracts visual isometric invariant features from normal-sample patches, motivated by industrial images’ rotations, translations, and flips.The paper proposes these features as a way to represent invariant information in the memory bank.Source: §2.2 · p. 4

2.3 GRAPH REPRESENTATION OF IMAGE

• Each normal image is split into patches whose feature vectors become graph nodes, with directed edges from each node’s K nearest neighbors.The resulting graph is G = (V, E).Source: §2.3 · p. 5

2.4 GRAPH FEATURE PROCESSING

• Graph feature processing uses GCN to exchange information between neighboring nodes, with max-relative graph convolution aggregating neighbor-feature differences.The aggregate and update operations use learnable weights; the aggregator max-pools feature differences and the updater is an MLP with batch normalization and ReLU.Source: §2.4 · p. 5

2.5 GRAPHCORE ARCHITECTURE

• GraphCore builds its memory bank with the Aug.(R) algorithm but uses VIIG without data augmentation as its feature extractor.Its test-time anomaly-score computation is otherwise highly similar to Augmentation+PatchCore.Source: §2.5 · p. 6

2.6 A UNIFIED VIEW OF AUGMENTATION+PATCHCORE AND GRAPHCORE

• The unified view presents PatchCore, Augmentation+PatchCore, and GraphCore as three architectures, with GraphCore using isometric invariant features.The paper links GraphCore’s feature choice to locating a feature subset for anomaly scoring.Source: §2.6 · p. 6, Table 1 · p. 6

3.1 EXPERIMENT SETTING

• The experiments compare GraphCore with unsupervised anomaly-detection methods reimplemented under the few-shot setting, alongside RegAD, which uses meta-learning across categories.The evaluation includes MVTec AD, MPDD, and MVTec LOCO AD; the reported average-AUC comparison covers MVTec AD and MPDD.Source: §3.1 · pp. 6–7

3.2 COMPARISON WITH THE SOTA METHODS

• GraphCore improves average AUC over SOTA models by 5.8%, 4.1%, 3.4%, and 1.6% on MVTec AD and 25.5%, 22.0%, 16.9%, and 14.1% on MPDD for 1-, 2-, 4-, and 8-shot cases.The comparison reports image and pixel AUROC; Table 2 covers average results across categories.Source: §3.2 · p. 7, Table 2 · p. 7
• On 2-shot results, GraphCore outperforms all baselines in 12 of 15 MVTec AD categories at image level and 11 of 15 at pixel level, plus 5 of 6 MPDD categories at image level and all at pixel level.The detailed category results are reported for 2-shot in the main text; other shot counts are in the appendix.Source: §3.2 · p. 7, Table 3 · p. 8, Table 4 · p. 8

3.3 ABLATION STUDIES

• Increasing the sampling rate from 0.0001 to 0.001 significantly improves performance, while further increases have a flattening effect.GraphCore’s performance is described as insensitive to sampling rate as it steadily increases.Source: §3.3 · p. 7, Figure 8 · p. 8
• Increasing nearest neighbors from 3 to 9 improves pixel-level performance at low sampling rates, but not image-level performance; the pixel-level gain approaches zero as sampling rises.Figure 8 identifies the 3-neighbor and 9-neighbor results by color.Source: §3.3 · pp. 7–8, Figure 8 · p. 8
• Rotation augmentation improves PatchCore on MVTec AD and MPDD, while GraphCore outperforms the augmented baseline by a large margin.The authors interpret the result as evidence for the effectiveness of rotation enhancement and isometric invariant feature extraction.Source: §3.3 · p. 8, Figure 9 · p. 9

3.4 VISUALIZATION

• In 1-shot visualizations across MVTec AD and MPDD, GraphCore produces satisfactory anomaly localization on various objects.The figure presents detection images, anomaly score maps, overlays, and ground truth; columns represent item types.Source: §3.4 · p. 9, Figure 10 · p. 9

4 CONCLUSION

• GraphCore captures isometric invariant features of normal industrial samples and outperforms SOTA models with at most eight normal training samples.The paper also presents Augmentation+PatchCore as a simple rotation-invariant feature baseline that significantly improves anomaly-detection performance.Source: §4 · p. 9

6.1 DATASET

• MVTec AD contains 15 item categories, while MPDD contains six metal-part classes and is described as more challenging because of varied viewpoints, distances, and nonuniform backgrounds.MVTec AD has 3,629 normal training images; MPDD has 888 normal training images.Source: §6.1 · p. 11
• For MVTec LOCO AD, the experiments retain normal and structural-abnormal test images but exclude logical-abnormal test images because their detection metric is calculated differently.The dataset also includes structural and logical abnormal images in its test set.Source: §6.1 · p. 11

6.2 EXPERIMENT RESULTS

• The cited appendix tables cover MVTec at 1-, 2-, and 4-shot with sampling ratio 0.01, reporting image AUROC and pixel AUROC across the listed few-shot settings.The captions specify the dataset, shot count, sampling ratio, and metric, but these passages do not provide cell values.Source: Table 5 · p. 12, Table 6 · p. 12, Table 7 · p. 12

6.3 ARCHITECTURE DETAILS

• GraphCore’s architecture details specify feature dimension, neighbor count, input size, and training configuration, including 300 epochs, AdamW, and cross-entropy loss.The architecture table identifies D as feature dimension, K as neighbor count, and H × W as input image size.Source: §6.3 · p. 17, Table 21 · p. 17

6.4 ABLATION STADIES

• The appendix reports that GraphCore’s memory-bank size and inference speed are effective in the 1- and 2-shot ablations, and that rotation outperforms other augmentation techniques.The augmentation comparisons are described for MVTec 2D at 1 and 2 shots with image-level AUROC.Source: §6.4 · p. 17, Table 22 · p. 17, Table 23 · p. 17