基于几何信息先验分布的似物性推荐方法

Objectness Proposal Based on Prior Distribution of Geometric Characteristics of Object Regions

似物性推荐是计算机视觉研究中的热门问题,其目的是用尽可能少的推荐窗口涵盖可能的兴趣目标,以显著地提升目标检测任务的计算效率。从组合几何学角度对该问题进行了分析,一种＂完全窗口覆盖＂的方法被提出,用少量窗口即可覆盖所有可能目标区域。对于尺寸不大于512×512的图像,约19000个窗口即可覆盖所有尺寸不小于16×16的目标区域。基于目标矩形的位置、尺寸的先验分布,可以使用贪心策略进一步地缩减窗口数量。为了适应不同图像集在小概率样本上的差异,提出了一种融合了贪心和随机方法的混合机制,其所需的计算量非常小,而且具有很好的泛化能力。在VOC2007测试集上,该混合机制可以在1000个推荐窗口上取得94.52%的召回率,其中在前10个热点推荐窗口上的召回率比其他方法平均高出13.99%~40.29%。

Objectness proposal is an emerging problem aiming to improve the efficiency of object detection by reducing candidate windows. The problem was analyzed from the perspective of combinatorial geometric , and a method was pro- posed to construct full cover sets which cover all possible object rectangles with a rather small amount of windows. For images no larger than 512 × 512, supposing all object rectangles are not smaller than 16 × 16, nearly 19000 windows are sufficient to make up a full cover set. By exploiting the prior distribution of locations/sizes of object rectangles, this amount can be reduced further in a greedy mode. In order to address the diversity of low-probability samples of different image sets, a hybrid scheme mixing the greedy and random methods which has good generality was presented. The new scheme recalls 94. 52% object rectangles with 1000 proposal windows, and its DRs on the first ten hot proposal win- dows are 13. 99%~40. 29% higher than existing methods in average.