Improving Robustness of Multiple-Objective Genetic Programming for Object Detection

Object detection in images is inherently imbalanced and prone to overfitting on the training set. This work investigates the use of a validation set and sampling rnethods in Multi-Objective Genetic Programming (MOGP) to improve the effectiveness and robustness of object detection in images. Results show that sampling methods decrease run-times substantially and increase robustness of detectors at higher detection rates, and that a combination of validation together with sampling improves upon a validation-only approach in effectiveness and efficiency.