## SceneSeg
Self-driving cars are usually trained to detect specific object types, such as cars, pedestrians, buses, etc. Such approaches are prone to failure cases when a self-driving car encounters an unusual object that it hasn't seen before, e.g. a rickshaw, or if a self-driving car encounters a strange presentation of a known object, e.g. a cyclist that has fallen over. In these scenarios, self-driving cars are unable to detect these critical objects leading to dangerous driving outcomes. To address this, we have developed SceneSeg, a neural network that is able to segment all important foreground objects, irrespective of what that object is. SceneSeg is able to implicitly learn the visual features of foreground objects such as cars, buses, vans, pedestrians, cyclists, animals, rickshaws, trucks and other similar objects, even though it has not been explicitly trained to detect these object types. SceneSeg is also able to detect objects that are outside of its training data, such …
