Pixel-Wise Anomaly Detection in Complex Driving Scenes

Figure 1 . Anomaly scenarios overview. There are three possible outcomes when a segmentation network encounters an anomalous instance. First, anomaly instances are properly segmented and classified as one of the training classes (i.e bird is confused as a person) (top). Second, anomaly instances are over-segmented with multiple classes (i.e dog is detected as a combination of person, vegetation, and terrain classes) (middle). And third, anomaly instances are blended with the background, not detected (i.e boxes blend with the street segmentation) (bottom). Our proposed method produces robust predictions for all scenarios, while previous approaches fail to handle at least one of them.