Revisiting Reverse Distillation for Anomaly Detection (CVPR 2023)

Code of CVPR 2023 paper: Revisiting Reverse Distillation for Anomaly Detection.

    The paper proposes the RD++ approach for anomaly detection by enriching feature compactness and suppressing anomalous signals through a multi-task learning design. For the feature compactness task, RD++ introduces the self-supervised optimal transport method. For the anomalous signal suppression task, RD++ simulates pseudo-abnormal samples with simplex noise and minimizes the reconstruction loss.
    RD++ achieves a new state-of-the-art benchmark on the challenging MVTec dataset for both anomaly detection and localization. More importantly, when compared to recent SOTA methods, RD++ runs 6.x times faster than PatchCore and 2.x times faster than CFA, while introducing a negligible latency compared to RD.

Table of Contents

• Libraries
• Data Preparations
• Train
• Evaluation
• Quick Experiments
• Citation
• Acknowledgement

Libraries

- geomloss
- numba

Data Preparations

Download MVTEC dataset from [Link]

Train

To train and test the RD++ method on 15 classes of MVTEC, for example, with two classes: carpet and leather, please run:

python main.py --save_folder RD++  \
               --classes carpet leather

Evaluation

If you only need to perform inference with checkpoints, please run:

python inference.py --checkpoint_folder RD++  \
                    --classes carpet leather

The pretrained weights can be found here [Google Drive]

Quick Experiments

Try the Colab here using

Citation

Please cite our paper if you find it's helpful in your work.

@InProceedings{Tien_2023_CVPR,
    author    = {Tien, Tran Dinh and Nguyen, Anh Tuan and Tran, Nguyen Hoang and Huy, Ta Duc and Duong, Soan T.M. and Nguyen, Chanh D. Tr. and Truong, Steven Q. H.},
    title     = {Revisiting Reverse Distillation for Anomaly Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {24511-24520}
}

Acknowledgement

We use RD as the baseline. Also, we use the Simplex Noise. We are thankful to their brilliant works!