Current and future trends in object detection

Abstract

The task of finding objects belonging to classes of interest in images has long been a focus of Computer Vision research. The ability to localize objects is useful in many applications: from self-driving cars, where it allows the car to detect pedestrians, bicyclists, road signs, and other vehicles, to security, where intruding persons can be detected. Though a lot of progress has been made since the conception of the field of Computer Vision more than five decades ago, as always, there is scope for further improvement. This is especially true in the case of object detection where a myriad of factors including variation in object instances through pose and appearance, along with other environmental factors such as the degree of occlusion, and lighting tend to cause failures. In this work we focus on improving object detection through the use of more representative features and better models. We propose new features that are not only more powerful, but also more robust and capture more information than the currently popular features. Further, we propose scalable models which can leverage large amounts of training data to improve performance.