On the Problem of Consistent Anomalies in Zero-Shot Industrial Anomaly Detection
Authors: Tai Le Gia, Jaehyun Ahn
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on MVTec AD using the Vi T-L-14-336 backbone show 98.3% AUROC for AC and AS performance of 66.8% (+4.2%) F1 and 68.1% (+5.4%) AP over state-of-the-art zero-shot methods. Additional experiments with the DINOv2 backbone further enhance segmentation, achieving a 69.1% (+6.5%) F1 and a 71.9% (+9.2%) AP, demonstrating the robustness of our approach across different architectures. |
| Researcher Affiliation | Academia | Tai Le-Gia EMAIL Department of Mathematics Chungnam National University Ahn Jaehyun EMAIL Department of Mathematics Chungnam National University |
| Pseudocode | Yes | Algorithm 1 Coverage-based Selection; Algorithm 2 Targeted Patch Filtering |
| Open Source Code | Yes | Our code is available at https://github.com/Dum Bringer/Co De Graph. |
| Open Datasets | Yes | We conducted experiments on two well-established benchmarks for industrial AC and AS: MVTec AD (Bergmann et al., 2019) and Visa (Zou et al., 2022). |
| Dataset Splits | Yes | In zero-shot anomaly detection and segmentation, we aim to identify defects in unlabeled test images D = {I1, . . . , IN} without any training data. We conducted experiments on two well-established benchmarks for industrial AC and AS: MVTec AD (Bergmann et al., 2019) and Visa (Zou et al., 2022). |
| Hardware Specification | Yes | All experiments use one RTX 4070Ti Super. |
| Software Dependencies | No | The paper mentions backbones like 'Vi T-L/14-336, pretrained by Open AI Radford et al. (2021)' and 'DINOv2 backbone', but does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | For the anomaly similarity graph, we selected the distance d(x, I(i)) to the i-th nearest image... The weighted endurance ratio was set with α = 0.2 and ω = 0.3 N. The coverage-based selection algorithm targeted a coverage of τ = 0.95. For anomaly scoring via the MSM in equation 2, we averaged the lowest 10% of distances... For the final anomaly scores, we used receptive field sizes r {1, 3, 5}. |