dlib-android | : dragon : Port dlib to Android | Android library

This is https://github.com/android-ndk/ndk/issues/82

The fix for this is to switch from gnustl to libc++, but note that libc++ isn't as stable as gnustl is yet (working on fixing this ASAP, should be ready by the time r15 hits stable).

EDIT: As of NDK r16 libc++ is the recommended STL. Switch to using libc++ (see our docs) for full C++11 (and beyond) support.

Looks like you need to change

Shape predictor trained mode file size depends on the count of points model training parameters. Existing model has the size of 98 MB and provide state-of-the-art accuracy of facial points detection. So you have three main ways to solve the problem:

1. some compression/quantization... that will change existing model. I will not help here - there are some answers in the question you are referenced
2. use different model training params. Existing model is trained with the code provided in "train_shape_predictor_ex" example in dlib package, but the training params are little different. simple running this example as-is with the images from example folder will give you model with the size of 15 MB. Yes, this model will be less-accurate, but this is the price of its small size
3. use less landmarks. to make this - use same "train_shape_predictor_ex" example, but remove some landmarks from training/testing dataset