Post-hoc feature-based out-of-distribution detection for real-world conditions
Samuel Wilson · Queensland University of Technology · 2024
This thesis addresses the challenge of detecting when vision models make erroneous, yet confident, predictions on new unknown inputs. The contributions proposed in this thesis detect these errors by monitoring the internal behaviour of the vision models during deployment, signalling when the model has behaved abnormally for a prediction. These contributions are researched in the context of mobile robot deployments in the real-world, emphasising the importance of computational efficiency, accuracy and risk reductions for the robot and anyone nearby.