DLIB：为 194 个地标训练 Shape_predictor（海伦数据集）

Question

我正在使用 helen dataset 为 194 个面部特征点训练 DLIB 的 shape_predictor，该数据集用于通过face_landmark_detection_ex.cpp 检测面部特征点dlib 库。

现在它给了我一个大约 45 MB 的 sp.dat 二进制文件，与 68 个面部标志的给定文件 (http://sourceforge.net/projects/dclib/files/dlib/v18.10/shape_predictor_68_face_landmarks.dat.bz2) 相比，它要小。训练中

• 平均训练误差：0.0203811
• 平均测试误差：0.0204511

当我使用经过训练的数据来获取面部地标位置时，我得到了结果..

与从 68 个地标得到的结果相差很大

68 地标图片：

为什么？

Answer 1

好的，看来你还没有读过code cmets (?)：

shape_predictor_trainer trainer;
// This algorithm has a bunch of parameters you can mess with.  The
// documentation for the shape_predictor_trainer explains all of them.
// You should also read Kazemi's paper which explains all the parameters
// in great detail.  However, here I'm just setting three of them
// differently than their default values.  I'm doing this because we
// have a very small dataset.  In particular, setting the oversampling
// to a high amount (300) effectively boosts the training set size, so
// that helps this example.
trainer.set_oversampling_amount(300);
// I'm also reducing the capacity of the model by explicitly increasing
// the regularization (making nu smaller) and by using trees with
// smaller depths.  
trainer.set_nu(0.05);
trainer.set_tree_depth(2);

看看Kazemi paper，ctrl-f 字符串'parameter'并读一读...