Avoiding over-detection: Towards combined object detection and counting

Abstract

Existing object detection frameworks in the deep learning field generally over-detect objects, and use non-maximum suppression (NMS) to filter out excess detections, leaving one bounding box per object. This works well so long as the ground-truth bounding boxes do not overlap heavily, as would be the case with objects that partially occlude each other, or are packed densely together. In these cases it would be beneficial, and more elegant, to have a fully end-to-end system that outputs the correct number of objects without requiring a separate NMS stage. In this paper we discuss the challenges involved in solving this problem, and demonstrate preliminary results from a prototype system.