Efficient and robust reduction of bounding boxes of a multi-class neural network’s output for vehicular radar-systems

Abstract: Object detection has been a fundamental part of many emerging technologies, such as autonomous vehicles, robotics, and security. As deep learning is the main reason behind the leap of performance in object detection, it has mostly been associated with a post-processing step of non-maximum suppression (NMS) to reduce the number of resulting bounding boxes output from the network to, ideally, one box per object. As non-maximum suppression blindly suppress the overlap with a pre-defined threshold, it introduces the problem of suppressing false negatives in crowded scenes by choosing a high threshold, or vice versa. This problem is critical, especially in the autonomous vehicle industry, as this concerns the safety of passengers. The problem of the machine understanding whether these bounding boxes belong to the same object or two near-by objects is still not directly solvable. Although a lot of previous research tried to invent a new box-reduction method, every method has its own drawbacks while solving the problem. That is why, until now, many researchers are still using non‐maximum suppression. In this research, a literature review was carried out to determine the best NMS alternatives. Then, an approach for box reduction based on determinantal point process (DPP) was implemented. Furthermore, an evaluation pipeline was introduced for experimental analysis for the differences between NMS and DPP. Although NMS shows a better performance in terms of precision and recall, DPP chooses better fitting bounding boxes.