Eliminating Useless Object Detectors Evolved in Multiple-Objective Genetic Programming

Abstract

Object detection is the task of correctly identifying and locating objects of interest within a larger image. An ideal object detector would maximise the number of correctly located objects and minimise the number of false-alarms. Previous work, following the traditional multiple-objective paradigm of finding Pareto-optimal tradeoffs between these objectives, suffers from an abundance of useless detectors that either detect nothing (but with no false-alarms) or mark every pixel as an object (perfect detection performance with but a very large number of false-alarms); these are very often Pareto-optimal and hence inadvertently rewarded. We propose and compare a number of improvements to eliminate useless detectors during evolution. The most successful improvements are generally more inefficient than the benchmark MOGP approach due to the often vast numbers of additional crossover and mutation operations required, but as a result the archive populations generally include a much higher number of Pareto-fronts.