One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection
Authors: Yiyue Li, Shaoting Zhang, Kang Li, Qicheng Lao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations across eleven datasets in three domains demonstrate our model s effectiveness compared to the latest AD methods. Additionally, our method has been proven to transfer flexibly to other AD methods, with the generated image data effectively improving the performance of other AD methods. 4 Experiments 4.2 Results 4.3 Ablation Study |
| Researcher Affiliation | Collaboration | Yiyue Li1 Shaoting Zhang4 Kang Li134 Qicheng Lao24 1West China Biomedical Big Data Center, West China Hospital, Sichuan University 2School of Artificial Intelligence, Beijing University of Posts and Telecommunications 3Sichuan University Pittsburgh Institute, Sichuan University 4Shanghai Artificial Intelligence Laboratory |
| Pseudocode | No | The paper describes its methods in text and provides an overview figure, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | Yes, the data in our paper is publicly available, and we will also make the code public in the future. |
| Open Datasets | Yes | We validate the effectiveness of our method across 11 datasets spanning three distinct domains: industrial, medical, and semantic domains. In the industrial domain, we utilize multiple datasets including MVTec-AD [2], Visa [46], KSDD [34], AFID [32], and ELPV [7]. For the medical domain, we incorporate datasets covering various modalities such as magnetic resonance imaging (MRI), computed tomography (CT), and optical coherence tomography (OCT). Specifically, the medical datasets include OCT2017 [17], Brain MRI, Head CT [13], and RESC [13]. For semantic anomaly detection, we employ two datasets: MNIST [19] and CIFAR10 [18]. |
| Dataset Splits | Yes | We follow [45] for the dataset partitions. We only use data from the MVTec-AD dataset as auxiliary training data. When testing on the MVTec-AD dataset, we use auxiliary data from the Visa dataset. Our anomaly detection task is a few-normal-shot setting including 2-shot, 4-shot, and 8-shot scenarios using only normal images. |
| Hardware Specification | Yes | All the experiments are trained by use of Py Torch on an NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | Yes | We select Stable Diffusion V1.5 [29] and utilize the CLIP with Vi T-L/14 architecture [28] for all tasks. Our anomaly-free customized model is fine-tuned using Dreambooth [31], and all model parameters are frozen in subsequent tasks. All the experiments are trained by use of Py Torch... |
| Experiment Setup | Yes | Our anomaly detection task is a few-normal-shot setting including 2-shot, 4-shot, and 8-shot scenarios using only normal images. The ratio of the t-step is set to 0.3. We set the parameters α and β for Ascore to 1 and 0.5, respectively, across all datasets. The image-level memory bank is set to 30, and the inference resolution is 240 240. |