December 2019
tl;dr: A new benchmark measuring how well methods detect potentially hazardous anomalies in driving scenes.
Embeddings of intermediate layers hold important information for anomaly detection.
- 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.
- Lost & Found dataset is real dataset, it can be used to compare the performance on synthetic datasets to identify methods that detect image in-painting instead of anomalies.
- Image inpainting postprocessing steps please refer to Augmented Reality Meets Computer Vision : Efficient Data Generation for Urban Driving Scenes IJCV 2018 (data augmentation with AR)
- evaluation scores: AP, and FPR@0.95Recall
- Foggy cityscape: dataset with adjustable visibilities.
- WildDash ECCV 2018 and RobustVisionChallenge CVPR 2018 for semantic and instance segmentation
- Safe Visual Navigation via Deep Learning and Novelty Detection (generative reconstruction)