Eyecandies

我们介绍了Eyecandies，这是一种用于无监督异常检测和定位的新型合成数据集。程序生成的糖果的照片逼真的图像在多个闪电条件下的受控环境中渲染，还提供了工业输送机场景中的深度和法线图。我们为模型训练和验证提供了无异常的样本，而仅在测试集中提供了具有精确地面真相注释的异常实例。该数据集包括十类糖果，每一类都显示出不同的挑战，例如复杂的纹理，自我遮挡和投机。此外，我们通过随机绘制过程渲染管道的关键参数来实现较大的类内变化，从而可以创建具有逼真外观的任意数量的实例。同样，将异常注入到渲染图中，并自动生成按像素排列的注释，从而克服了人为偏见和可能的不一致。

We introduce Eyecandies, a novel synthetic dataset for unsupervised anomaly detection and localization. Photorealistic images of procedurally generated candies are rendered in controlled environments with multiple lighting conditions, alongside depth and normal maps from industrial conveyor belt scenarios. We provide anomaly-free samples for model training and validation, while anomaly instances with precise ground-truth annotations are only included in the test set. The dataset includes ten categories of candies, each presenting distinct challenges such as complex textures, self-occlusion, and specular highlights. Furthermore, we achieve large intra-class variations by randomly sampling key parameters of the rendering pipeline, enabling the creation of an arbitrary number of instances with photorealistic appearances. Similarly, anomalies are injected into the rendered images, and pixel-wise annotations are automatically generated, thereby eliminating human bias and potential inconsistencies.