Oriented Object Proposals

ICCV 2015

关于候选区域提取的文献。 Oriented Object Proposals 简写 OOP算法
本文提出的提取算法的三个优势：
1）可以提取不同旋转角度的物体
2）本文的OOP算法得到更好的框，框的更准更小些
3）我们的候选区域个数更小些

2 Related Work
以前候选区域提取算法主要可以分为两类：
1）Window scoring methods 滑动窗口打分方法
2）Superpixel grouping methods 聚类方法

其实还有一类 深度学习方法：文献 Faster R-CNN 中提取方法

本文算法简要：
We formulate the object proposal generation problem as a generative probabilistic model, where the like-lihood function corresponds to how likely a pixel belongs to an object. In order to describe the objectness of a pixel without bounding boxes, we introduce a background prior, together with a gradient map and structured edges [9], to construct the likelihood function. The objects are searched and located as the local maximum log-likelihood and the shape of the object is described by a full covariance matrix. The optimum position and shape are obtained by an efficient iterative searching algorithm.

3.1.1 Background Prior
这里我们基于一个假设：图像边缘附近的像素不属于显著物体

通过定义一个像素度函数，通过迭代优化候选区域位置和形状得到最后的候选区域。

有点计算 objectness 的意思

算法的时间：
We have implemented it in C++ and tested it on a PC with an i7 3.4GHz CPU and 8GB RAM. It takes on average 0.5s to process a 500×400 image, with the iteration process taking about 0.4s

没用GPU