December 2019
tl;dr: A new benchmark measuring how well methods detect potentially hazardous anomalies in driving scenes.
Overall impression
Embeddings of intermediate layers hold important information for anomaly detection.
Key ideas
- Bayesian DL: epistemic uncertainty, aleatoric uncertainty, distributional uncertainty
- Novelty detection (Out of distribution detection): one class cls which aim at discriminative embeddings, density estimations, and generative reconstruction.
- Softmax score is not a reliable score for anomaly detection
- Most better performing methods require special loss that reduced segmentation accuracy (tradeoff between better outlier detection and error. Cf tradeoff between better uncertainty calib and error)
- Learning anomaly detection from fixed OoD data is on par with unsupervised methods for most of the datasets. Void classifier is most practical way forward. A separate void class is concisely better than maximizing the softmax entropy. A separate void class is also most practical.
Technical details
Notes