As humans, we sometimes know that we know too little to make a decision. We need to develop similar uncertainty in computer systems.
Not long ago, we saw the first fatal accident caused by an autonomous car. In May 2016, a perception system confused the white side of a trailer for a bright sky. The car drove into the trailer and killed the person in the front seat. Currently, the perception systems used in autonomous cars are built on low-level features that are blindly assumed to be correct, which is not always the case. These low-level features are extracted without quantifying their associated uncertainties. This means that erroneous predictions, such as identifying the white side of the train as a bright sky, are propagated through the system and can lead to fatal decisions (like driving into the trailer). If the perception system had been able to assign a high level of uncertainty to its erroneous prediction, this accident might have been avoided.
In this project, we seek to quantify the system uncertainty and try to understand what the system does not know. We focus on visual place recognition (VPR), which is the task for retrieving an exact location only using images. This is an important low-level feature for the long-term operation of autonomous vehicles, and is challenging in practical settings due to appearance variability between two views of a scene taken at different times and from different viewpoints.
We combat the lack of uncertainty estimates with Bayesian Deep Learning, i.e. we model the VPR problem with a Bayesian Neural Network where the model has a prior distribution over its weights and use data to learn its posterior distribution. Given a new place, we aim to retrieve images from the same location with accurately calibrated uncertainty estimates. As a consequence, we hope to understand when the model knows too little to make a decision, and thus alleviate propagation of erroneous low-level features, which could prevent fatal accidents caused by autonomous vehicles.