Improving Region Based CNN Object Detector Using Bayesian Optimization

With impressive results on many object detection benchmarks, Deep Neural Networks (DNN) proved to be revolutionary technology when it comes to object detection task. One major advantage of DNN is the high capacity and strong differentiation ability of the trained models, still imprecise localization is an important error source. In this paper, we present a sequential searching algorithm using Bayesian Optimization to propose better bounding boxes hence reducing localization error.We tackle the localization difficulty by transforming region proposal into an optimization problem and using Bayesian Optimization as a black-box optimizer to iteratively solve it. The proposed algorithm demonstrated the state-of-the-art performance on PASCAL VOC 2007 benchmark under the standard localization requirements.