A Unified Model for Multi-class Anomaly Detection

Authors: Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, Xinyi Le

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our algorithm on MVTec-AD and CIFAR-10 datasets, where we surpass the state-of-the-art alternatives by a sufficiently large margin.
Researcher Affiliation Collaboration 1Shanghai Jiao Tong University 2Tsinghua University 3CUHK 4Sense Time
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. The method is described in text and via diagrams.
Open Source Code Yes Code is available at https:// github.com/zhiyuanyou/Uni AD.
Open Datasets Yes MVTec-AD [4] is a comprehensive, multi-object, multi-defect industrial anomaly detection dataset with 15 classes. CIFAR-10 [23] is a classical image classification dataset with 10 categories.
Dataset Splits No The paper describes how classes are used for training and testing (e.g., normal samples for training, anomalous for testing, unified/separate cases for MVTec-AD and CIFAR-10 class combinations), but does not explicitly provide percentages or counts for train/validation/test dataset splits, nor does it clearly define a separate validation set with specific size or methodology.
Hardware Specification Yes Our model is trained for 1000 epochs on 8 GPUs (NVIDIA Tesla V100 16GB) with batch size 64.
Software Dependencies No The paper states 'implemented in Python 3.8' but does not specify versions for other key software components or deep learning frameworks used (e.g., PyTorch, TensorFlow, or specific library versions).
Experiment Setup Yes The image size is selected as 224 224, and the size for resizing feature maps is set as 14 14. The reduced channel dimension is set as 256. Adam W optimizer [21] with weight decay 1 10 4 is used. Our model is trained for 1000 epochs on 8 GPUs (NVIDIA Tesla V100 16GB) with batch size 64. The learning rate is 1 10 4 initially, and dropped by 0.1 after 800 epochs. The layer numbers of the encoder and decoder are both 4. The neighbor size, jittering scale, and jittering probability are set as 7 7, 20, and 1, respectively. The evaluation is run with 5 random seeds.