注: 刚刚看完了这篇论文,顺便整理了一下这篇论文的思路,对论文的方法表示666
项目主页:http://www.cbsr.ia.ac.cn/users/scliao/projects/lomo_xqda/
论文解析
- 行人重识别有两个特别重要的问题:
- Feature representation
- Metric learning
在这篇论文中,作者采用的是LOMO的特征表示方法,以及XQDA的度量学习方法
LOMO:Local Maximal Occurrence Representation, 工作原理如下:
- The LOMO feature analyze the horizontal occurrence of local features, and maximizes the occurrence to make a stable representation against viewpoint change. Besides, to handle illumination variations, we apply the Retinex transform and a scale invariant texture operator.
- 首先,使用Retinex algorithm 对图片进行预处理,减少光照对重识别带来的影响
- Retinex considers human lightness and color perception. It aims at producing a color image that is consistent to human observation of the scene
从效果图中可以看到,在a中不同view下illumination的差异在b中基本被消除了,这一定程度上减少了光照对重识别带来的影响
- 接着,对使用Retinex algorithm得到的Retinex image使用HSV color historgram提取颜色特征
- 除了color description, 作者还采用Scale Invariant Local Ternary Pattern(SILTP)的方法获得illuminaton invariant texture description
- 以上主要解决illumination change的问题,接下来要解决viewpoint change的问题
- 在之前的工作中,有论文proposed:equally divided a person image into six horizontal stripe, and a single histogram is comouted in each stripe.
- 虽然这种方法在一定程度上可以减轻viewpoint change带来的影响,但是这种方法可能会导致空间细节的丢失,进而影响其判别的能力
- 因此,作者提出使用silding windows来完成描述的工作:
- Specifically, we use a subwindow size of 10X10, with an overlapping step of 5 pixels to locate local patches in 128X48 images. Within each subwindow, we extract two scale of SILTP histograms, and an 8X8X8-bin joint HSV historgam. Each histogram bin represents the occurence probability of one pattern in a