CVPRW 2022

论文名称：Anomaly Detection in Autonomous Driving: A Survey

corner case 定义为，检测难度由易到难，本文只关注自然场景下的 corner case，所以 pixel-level 的 case 不在本文讨论范围内：

• pixel
• domain
• object
• scene
• scenario

检测 corner case 的方式有以下几种：

• Confidence score techniques are often derived by post-processing without interfering with the training of a neural network and subdivided into Bayesian approaches, learned scores, and scores obtained by post-processing.
• Reconstructive approaches try to reconstruct normality and consider any kind of deviation from it as anomalous.
• Generative approaches are closely related to the former reconstructive approaches, but also take into account the discriminator’s decision or the distance to the training data.
• Feature extraction can be based on handcrafted or learned features to determine a class label or compare modalities on various feature levels. Prediction based techniques predict the next frame(s) expected under normality.