Unveiling Multi-View Anomaly Detection: Intra-view Decoupling and Inter-view Fusion

Authors: Kai Mao, Yiyang Lian, Yangyang Wang, Meiqin Liu, Nanning Zheng, Ping Wei

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed IDIF achieves state-of-the-art performance on three datasets. Extensive ablation studies also demonstrate the superiority of our methods.
Researcher Affiliation Academia National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University, Xi an, China Ping Wei is the corresponding author (EMAIL).
Pseudocode No The paper describes the methodology using textual explanations, mathematical equations, and figures (Figure 2 illustrates the overall framework, and Figure 3 illustrates View-wise Dropout). However, it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/Kerio99/IDIF
Open Datasets Yes We evaluate our method on the large-scale, multi-view anomaly detection dataset Real-IAD (Wang et al. 2024). We additionally process two 3D datasets Mv Tec 3D-AD (Bergmann et al. 2019) and Eyecandies (Bonfiglioli et al. 2022) into the multi-view 2D form.
Dataset Splits No The paper mentions training and testing phases and refers to Xtrain and Xtest, but does not provide specific percentages, counts, or methodologies for how these splits are created or used, nor does it cite standard splits with precise details for reproducibility.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions using ResNet18 as a teacher network.
Software Dependencies No The paper mentions using ResNet18 and ImageNet for pretraining, and Describable Textures Dataset (DTD) as the source of pseudo anomalies, but it does not specify any version numbers for programming languages, libraries, or other software components.
Experiment Setup Yes Images are resized to 256 256 for training and testing. The batch size is 8. We use the Describable Textures Dataset (DTD) (Cimpoi et al. 2014) as the source of pseudo anomalies. With the dropout probability of each view set to 0.2 during training, the expected number of times for each combination of four missing views is less than twice in 1000 steps of training, while the strategy helps preserving 94.0% S-AUROC performance.