Spatial Non-Maximum Suppression for Object Detection using Correlation and Dynamic Thresholds

This paper presents a spatial-non-maximal suppression algorithm (SNMS) that is hardware friendly and improves efficiency. Unlike the greedy NMS, which merely focuses on the scores, the SNMS considers the correlation among the overlapped boxes and applies dynamic thresholds which are determined by the boxes density. Thus, the SNMS integrates spatial and contextual information with the scores. Three techniques are used to shorten the latency. First of all, the SNMS starts the suppression as soon as the 1
^st
candidate box is ready. Secondly, the candidate boxes on different layers are analyzed in parallel. Thirdly, the representatives of boxes clusters (BC) which consist of overlapped boxes at multiply scales, are selected on the fly based on the forecastable position relationships. Experimental results demonstrate that the SNMS can compress the number of candidate boxes 30 times at maximum and maintains the accuracy at the same time.