我觉得你可以看一下这篇文章，“Towards Practical Facial Feature Detection”，
有关样本数及区域选择等因素进行了非常深入的研究分析，而且是07年的文章，也算比较新吧。其中提到：
We investigated the effects of the training set on performance. The number of
positive examples was varied from 100, to 20000 according to the following schedule :
{100; 300; 500; 1000; 2000; 3000; 5000; 8000; 10000; 15000; 20000}. For each condition the number of negative examples was 3 times the number of positive examples.
另外，不知可否将你的训练样本及描述文件（info等）发一份给我？sweetjujube@sina.com
发表于 2008-12-29 17:18:21
