A simplified non-maximum suppression with improved constraints

Post processing plays an important role in object detection methods based on the sliding window method. Simplified non-maximum suppression is a typical representative of post processing methods. However, traditional simplified non-maximum suppression uses only one constraint and cannot discard repetitive detections effectively. An improved simplified non-maximum suppression with two additional constraints was proposed. Compared with the traditional simplified non-maximum suppression which only calculates the proportion of intersection area to that of candidate detection bounding box, the two additional constraints named "completely covered detection suppression" and "PASCAL VOC overlap criterion" calculate the proportions of the intersection area to that of the selected detection bounding box and to the union area, respectively. The experimental results show that the improved simplified non-maximum suppression could discard the false positives effectively and significantly improve detection performance.