Ensemble Methods for Object Detection
Ángela Casado-García, Heras Jónathan
Abstract
Object detection is one of the most important topics of computer vision since it has many applications in several fields. Object detection models can be improved thanks to ensemble techniques; however, the process of ensembling object detectors poses several challenges. In this paper, we present an ensemble algorithm that can be applied with any object detection model independently of the underlying algorithm. In addition, our ensemble method has been employed to define a test-time augmentation procedure for object detection models. Our ensemble algorithm and test-time augmentation procedure can be used to apply data and model distillation for object detection, two semi-supervised learning techniques that reduce the number of necessary annotated images to train a model. We have tested our methods with several datasets and algorithms, obtaining up to a 10% improvement from the base models. All the methods are implemented in an open-source library.