What, where and how many? Combining object detectors and CRFs
Ladický, L'Ubor (Oxford Brookes University, United Kingdom); Sturgess, Paul; Alahari, Karteek; Russell, Chris; Torr, Philip H. S. Source: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), v 6314 LNCS, n PART 4, p 424-437, 2010, Computer Vision, ECCV 2010 - 11th European Conference on Computer Vision, Proceedings
Database: Compendex
Abstract  -  Detailed

Generalization: This paper define a globe energey function for the model, which combines results from sliding window detectors, and low-level pixel-based unary and pairwis relations.

motivation: try to recognition objects, find their location and spatial extent, and also provide the number of instances of objects. This work can be viewed as an integration of object class segmentation methods,which fail to distinguish between adjacent instances of objects of the same class, and object detection approaches, which do not provide information about background classes.

创新点：定义了一个全局能量函数 = object detector + pairwise + unary, 并能够高效求解。

       pairwise: 利用mid-level cues 之间的关系，如superpixels之间的关系

       unary:利用low-level cues, 即pixel-based

       object detector:

能量函数：

,其中后一项即为检测的能量项，前一项为基于像素或超像素的单位置或双位置势函数。

其中f为假设函数，g为惩罚函数

最好检测项归结为：

该函数形式具有Robust ，因此可以用和算法高效地实现。

实验结果：第二列为不加检测的，第三类为加了检测的。表现在加了检测的对象轮廓更加完整