Fusion ensemble methods to improve object detection accuracy

Object detection finds place in real-life applications in a variety of domains ranging from agriculture to defense. Increasing detection accuracy is a step toward the use of deep learning based object detectors in safety-critical applications. We propose increasing the accuracy using statistical fusion ensemble method on the bounding box coordinates obtained from multiple object detection models, using root mean squared error as a metric. Tests on the PascalVOC 2012 dataset indicate that the statistical fusion ensemble method is optimal for 40% of the test cases, implying that the individual detectors collectively fail by the same percentage.