Non Maximal Suppression in Cascaded Ranking Models

Ranking models have recently been proposed for cascaded object detection, and have been shown to improve over regression or binary classification in this setting [1, 2]. Rather than train a classifier in a binary setting and interpret the function post hoc as a ranking objective, these approaches directly optimize regularized risk objectives that seek to score highest the windows that most closely match the ground truth. In this work, we evaluate the effect of non-maximal suppression (NMS) on the cascade architecture, showing that this step is essential for high performance. Furthermore, we demonstrate that non-maximal suppression has a significant effect on the tradeoff between recall different points on the overlap-recall curve. We further develop additional objectness features at low computational cost that improve performance on the category independent object detection task introduced by Alexe et al. [3]. We show empirically on the PASCAL VOC dataset that a simple and efficient NMS strategy yields better results in a typical cascaded detection architecture than the previous state of the art [4, 1]. This demonstrates that NMS, an often ignored stage in the detection pipeline, can be a dominating factor in the performance of detection systems.