Unlocking the Potential of Reverse Distillation for Anomaly Detection
Authors: Xinyue Liu, Jianyuan Wang, Biao Leng, Shuo Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several benchmarks prove that our method outperforms existing unsupervised AD methods under RD paradigm, fully unlocking RD s potential. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Beihang University 2School of Intelligence Science and Technology, University of Science and Technology Beijing 3Beijing Key Lab of Traffic Data Analysis and Mining, School of Computer & Technology, Beijing Jiaotong University EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using textual explanations and diagrams (Figure 2, Figure 3, Figure 4) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/hito2448/URD |
| Open Datasets | Yes | We conduct our experiments primarily on MVTec AD (Bergmann et al. 2019) containing 5354 images across 15 categories, MPDD (Jezek et al. 2021) containing 1346 images across 6 categories, and BTAD (Mishra et al. 2021), which includes 2540 images across 3 categories. All datasets have only normal images in the training set, while have both normal and anomalous images in the test set. |
| Dataset Splits | Yes | All datasets have only normal images in the training set, while have both normal and anomalous images in the test set. |
| Hardware Specification | No | The paper does not provide specific hardware details (GPU models, CPU types, memory) used for running the experiments. It mentions parameters like 'training batch size is 16' but not the hardware. |
| Software Dependencies | No | The paper mentions using an 'Adam optimizer' and a 'Gaussian filter' and a 'Wide ResNet50' pretrained on 'ImageNet', but it does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific library versions). |
| Experiment Setup | Yes | During both training and inference, all images are resized to 256 256. The training batch size is 16, with an early stopping strategy for a maximum of 10k iterations. Consistent with RD, the student decoder is optimized using an Adam optimizer with a learning rate of 0.005, while the teacher encoder is trained with another one at a learning rate of 0.0001. During inference, the anomaly maps are smoothed using a Gaussian filter with σ = 4. |