The MVTec 3D-AD Dataset for Unsupervised 3D Anomaly Detection and Localization#

  • Authors: Paul Bergmann, Xin Jin, David Sattlegger, Carsten Steger

  • Affiliations: MVTec Software GmbH, Technical University of Munich, Karsruhe Institute of Technology

  • Links: Project Page, arXiv

Summary#

The authors introduce the first comprehensive 3D dataset for the task of unsupervised anomaly detection and localization. A high-resolution industrial 3D sensor is used to acquire depth scans of 10 different object categories.

../_images/fig-014.png

Key Ideas#

The MVTec 3D-AD dataset. The dataset consists of 4147 scans acquired by a high-resolution industrial 3D sensor. For each category, a set of anomaly-free scans is provided for model training and validation.

Performance evaluation. The model output a real-valued anomaly score for each \((x, y, z)\) pixel of the test set, which is then converted to binary predictions using a threshold. Then we compute the per-region overlap (PRO) metric, which is defined as the average relative overlap of the binary prediction \(P\) with each connected component \(C_k\) of the ground truth.

\[\text{PRO} = \frac{1}{K} \sum_{k=1}^K \frac{\lvert P \cap C_k \rvert}{\lvert C_k \rvert}\]

This process is repeated for multiple thresholds and a curve is constructed by plotting the resulting PRO values against the corresponding false positive rates. The final performance is computed by integrating this curve up to a limited false positive rate (i.e., 0.3) and normalising the resulting area to the interval \([0, 1]\).

Initial benchmark. Very few methods have been proposed explicitly for this task and all of them operate on voxel data. The authors consider Voxel f-AnoGAN [1] and their implementation of a convolutional Voxel AE [2].

Technical Details#

Visualization of the provided data for one anomalous test sample of the dataset category peach.

../_images/fig-024.png

Anomaly localization results.

../_images/fig-032.png

Notes#

References#

[1] Simarro Viana, Jaime, Ezequiel de la Rosa, Thijs Vande Vyvere, David Robben, and Diana M. Sima. “Unsupervised 3d brain anomaly detection.” In International MICCAI Brainlesion Workshop, pp. 133-142. Springer, Cham, 2020.

[2] Bengs, Marcel, Finn Behrendt, Julia Krüger, Roland Opfer, and Alexander Schlaefer. “Three-dimensional deep learning with spatial erasing for unsupervised anomaly segmentation in brain MRI.” International journal of computer assisted radiology and surgery 16, no. 9 (2021): 1413-1423.