Object Detection for automotive, aerospace, medical, or industrial applications relies heavily on Deep Neural Networks. Until 2012 – with classical algorithms – it was not possible to distinguish a child on the street from a plastic bag. Today, Deep Neural Networks can learn the difference between the plastic bag and the child by means of Deep Learning.
Neural Networks, however, are not always reliable, e.g. when the present object is not similar to the objects in the training data. We call these objects anomalies. Neural Networks do not have a concept of anomalies and will misclassify the unknown objects.
Let’s watch a movie to see what this is all about.
Level Five Safety develops technology to detect anomalies. Our product Avant[‘guard’] enables Neural Networks to distinguish the types of objects known and the types of objects that have never been seen during training. Avant[‘guard’] provides dependable confidence estimates for the objects detected.
Our technology evaluates the similarity of the object to the training data by means of mathematical-statistical methods. Our technology extends the architecture and the learning algorithm of a Neural Network by adapted and newly developed elements, which in combination allow a verifiable, superordinate classification into the groups known and unknown.
Our technology can be integrated with most Deep Learning applications, e.g. 2D Semantic Segmentation of images for Object Detection and Localization, Single-Shot Object Detection, and 3D Semantic Segmentation of Point Clouds. Our technology is complementary to the learned functionality of the Neural Network, so that the learned functionality remains unchanged while Avant[‘guard’] implements the verifiable and certifiable detection of unknown situations for the Neural Network.