该文的作者是Pedro F. Felzenszwalb等,著名的DPM目标检测模型的作者,该文的工作是将DPM(可形变组件模型)级联化以加快检测速度。加快的方式就是依次计算到目前为止部分部件得分之和,如果小于一定的阈值,则放弃该位置对象的继续检测,就是级联的思想。
作者在摘要中说到该文的一个核心贡献:
In analogy to probably approximately correct (PAC) learning, we introduce the notion of probably approximately admissible (PAA) thresholds. Such thresholds provide theoretical guarantees on the performance of the cascade method and can be computed from a
small sample of positive examples.
大意是提出了“概率近似容许(PAA)阈值的概念,该阈值在性能上有理论保证,并且只需要少量(训练)正样本。
在级联分类器中,各级阈值的设定是很重要的,有方法能确定有理论保证的阈值是非常好的。
看看该文中对阈值的设定方法和关于理论保证的证明吧,
先是一些概念说明:
We say a sequence of thresholds t is (; ) probably approximately admissible (PAA) if the probability that the error of t is greater than  is bounded by ,
Let be the optimal displacements for the nonroot parts of M on an example. We can define partial scores that take into account the first i parts and the first i parts minus the i-th deformation cost,
然后是阈值设置方法:
The star-cascade algorithm will find an optimal configuration and score for x if and only if and for all .
Let X be m independent samples from D. We can select thresholds by picking,
然后是理论证明:
我的疑问是:作者声称他们的阈值具有理论保证,但是定理1的证明是假设了阈值能保证误差很小。这不关键的结论-阈值有理论上的性能保证-没有得到证明吗?