Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)

Multi-shot Pedestrian Re-identification via Sequential Decision Making

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第1张图片

因为视频连续帧有很多冗余,解决方法:

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第2张图片

一个流学习rgb表面特征,一个流学习运动信息

他们的方法:

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第3张图片

出发点:效果好的时候,看一张图片就好了,出现模糊和遮挡的现象时,需要动态的调节看多少图像。

使用增强学习框架

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第4张图片

Then, the agent could output one of three actions: same, different or unsure

gt=groundtruth

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第5张图片

State部分:

1.single-shot网络提出的特征

2.之前状态特征的均值

3.手动指定的特征

论文解释:还通过手工特征增强了图像功能,以便更好地进行区分。对于每个时间步t,我们计算距离和内积f(xi)·f(yj) 对所有1≤i,j≤t - 1,然后将它们的最大值,最小值和平均值加到输入中,从而得到6维额外特征。

 

Unsure 权值决定看几张图。

 

优点:只使用4张图像就能很快超过使用全部200张的图片。

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第6张图片

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第7张图片

遮挡时候unsure的Q value要大!

 

State-of-the-art的比较数据

Multi-shot Pedestrian Re-identification via Sequential Decision Making (note)_第8张图片

 

样本时随机sample,按顺序选择样本效果要差一点。

你可能感兴趣的:(行人重识别)