Small object detection using retinanet with hybrid anchor box hyper tuning using interface of Bayesian mathematics

Abstract

In recent years object detection system has been improved by many folds due to many novel deep learning models. Deep learning has outperformed the existing traditional computer vision techniques. In recent many deep learning models uses the concept of anchor box, the model proposes various anchor boxes on the images. The models generally use a classification model and a regression models, the regression model is used to predict the position of next possible anchor box and the classification is used to validate the anchor box. The hyper tuning of these models are generally based on the anchor box specifications, many researchers have used an optimized anchor box dimensions which is obtained for a specific dataset, due to which the accuracy increases drastically but the model are not scalable on any other data set. We propose a new hybrid anchor box optimization technique by using a variant of Bayesian optimization and sub sampling for small object detection using retina net model with resnet backbone. Our hybrid model is scalable over various datasets, the model is used on visdrone dataset and the result shows a 3.7% improvement in MAP result.