SoftPatch: Unsupervised Anomaly Detection with Noisy Data
Authors: Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai WU, Yong Liu, Chengjie Wang, Feng Zheng
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments in various noise scenes demonstrate that Soft Patch outperforms the state-of-the-art AD methods on the MVTec AD and BTAD benchmarks and is comparable to those methods under the setting without noise. |
| Researcher Affiliation | Collaboration | 1Southern University of Science and Technology, Department of Computer Science and Engineering 2Tencent, Youtu Lab |
| Pseudocode | No | The paper includes diagrams (e.g., Figure 1, Figure 2) to illustrate the method but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code can be found in https://github.com/TencentYoutuResearch/Anomaly Detection-Soft Patch. |
| Open Datasets | Yes | Datasets. Our experiments are mainly conducted on the MVTec AD and BTAD benchmarks[11; 12]. MVTec AD contains 15 categories with 3629 training images and 1725 test images in total |
| Dataset Splits | Yes | MVTec AD contains 15 categories with 3629 training images and 1725 test images in total |
| Hardware Specification | Yes | All our experiments are run on Nvidia V100 GPU and repeated three times to report the average results. |
| Software Dependencies | No | The paper mentions using Wide-Res Net50 as backbone and other methods like Patch Core, Pa Dim, CFLOW, but does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | In the absence of specific instructions, the backbone of the feature extractor is Wide-Res Net50, and the coreset sampling ratio of Patch Core and Soft Patch is 10%. For MVTec AD images, we only use 256 x 256 resolution and center crops them into 224 x 224 along with a normalization. For BTAD, we use 512 x 512 resolution. We train a separate model for each class. ... we set the threshold τ in Soft Patch and the LOF-K to constant 0.15 and 6 for all noisy scenarios and classes. |