学术论文

      似物性窗口融合方法比较研究

      A comparative study of box aggregation methods for objectness proposals

      摘要:
      为研究似物性采样窗口融合方法对定位精度、召回率、采样候选窗口数量的影响,对采样后的窗口进行线性加权、非线性加权、密集区域融合.通过在PASCAL VOC 2007测试数据集上实验,绘出召回率-交并比、召回率-采样数、平均召回率评价曲线.研究结果表明:融合后可以显著降低窗口数量,同时保持较高的定位精度,在准确的似物性得分帮助下,似物性加权表现更优.
      Abstract:
      In order to study the effect of box aggregation for object proposals on the positioning location accuracy, recall, as well as candidate bounding boxes, a few box aggregation approaches including linear weighting, nonlinear weighting and intensive regional merging arestudied.The experiments on PASCAL VOC 2007 dataset have been conducted for evaluating the quality of object proposals, including recall-IoU, recall-proposal, and average recall curve.The experimental results show that box aggregation can reduce the number of candidate bounding boxes, while maintaining a high location accuracy, and objectness-weighting aggregation gets more favorable performance with the help of accurate objectnessscore.
      Author: LI Jin-dong CHEN Shu-han HU Xue-long
      作者单位: 扬州大学信息工程学院,江苏扬州,225127
      年,卷(期): 2017, 42(2)
      分类号: TP37
      在线出版日期: 2017年5月16日
      基金项目: 江苏省自然科学基金资助项目