摘要
尽管在图像异常检测和分割方面取得了显著进展，但很少有方法利用三维信息。我们利用最近引入的三维异常检测数据集来评估使用三维信息是否是一个错失的机会。首先，我们呈现了一个令人惊讶的发现：仅使用颜色的标准方法在性能上超过了所有当前专门设计用于利用三维信息的方法。这一结果违反直觉，因为即使是对数据集进行简单的检查也表明，对于包含几何异常的图像，仅使用颜色的方法是不够的。这促使我们提出一个问题：如何使异常检测方法有效利用三维信息？我们研究了一系列形状表示方法，包括手工设计和基于深度学习的方法；实验结果表明，旋转不变性在性能中起着主导作用。我们揭示了一种简单的仅使用三维信息的方法，该方法在不依赖深度学习、外部预训练数据集或颜色信息的情况下，超越了所有近期的方法。由于仅使用三维信息的方法无法检测颜色和纹理异常，我们将该方法与基于颜色的特征相结合，显著优于以往的最佳方法。我们的方法被命名为BTF（Back to the Feature），在MVTec 3D-AD数据集上实现了像素级ROCAUC：99.3%和PRO：96.4%的成绩。
代码仓库
eliahuhorwitz/3D-ADS
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 3d-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (FPFH) | Detection AUROC: 0.782 Segmentation AUPRO: 0.924 Segmentation AUROC: 0.978 |
| depth-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (Depth iNet) | Detection AUROC: 0.675 Segmentation AUPRO: 0.755 Segmentation AUROC: 0.930 |
| depth-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (SIFT) | Detection AUROC: 0.727 Segmentation AUPRO: 0.910 Segmentation AUROC: 0.974 |
| depth-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (RaW) | Detection AUROC: 0.573 Segmentation AUPRO: 0.442 Segmentation AUROC: 0.771 |
| depth-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (NSA) | Detection AUROC: 0.696 Segmentation AUPRO: 0.5572 Segmentation AUROC: 0.817 |
| depth-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (HoG) | Detection AUROC: 0.559 Segmentation AUPRO: 0.771 Segmentation AUROC: 0.930 |
| rgb-3d-anomaly-detection-and-segmentation-on | Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection (BTF) | Detection AUCROC: 0.865 Segmentation AUCROC: 0.992 Segmentation AUPRO: 0.959 |