Increasing Trustworthiness of Edge AI by Adding Uncertainty Estimation to Object Detection

Object detection is a common machine learning task that can be used to detect various categories of entities, e.g., people, in images and mark them with bounding boxes and a classification score.

Still, typical models are not able to fully express the uncertainty in their predictions, hence it is usually not known when a model prediction is trustworthy or not. However, a model with uncertainty estimation is able to say ”I don't know“, when it encounters an input unlike anything seen before during training.

Challenges of uncertainty estimation typically are bigger model sizes and higher inference times, which are problematic for real-time applications at the resource-constrained Edge.

The goal of this thesis is to implement and evaluate different approaches of uncertainty estimation for object detection. Evaluation checks how well the model distinguishes between input data it has been trained to process, and out-of-distribution samples it does not have the capabilities to handle well.