How to Use Ceres-Solver Library for Android-NDK

After using Levenberg-Marquardt optimization library from Lourakis in Android-NDK, this time I tried another library that offers non-linear least square solver called Ceres-Solver Library from Google.

https://code.google.com/p/ceres-solver/


From the specification and the demo, it is clear that this library shows impressive result (google map demo). And most importantly, this library already put one folder JNI inside it. It means we can quickly use it for our Android application.

For augmented reality application, I use ceres-solver for minimizing the reprojection error of the camera pose that is produced by our algorithm (refer this work for the detail). The result is comparable with the result I produced in the last post using Levenberg-Marquardt library. The good point is this solver offers sparse solver as well, and as written in the tutorial, it can be applied for solving the bundle adjustment step, which is suitable for our purpose.

The step for using ceres-solver for android-ndk:

1. download the library from https://code.google.com/p/ceres-solver/ (eigen library is necessary as well)
2. go to the jni folder after extracting it
3. build the library using ndk-build. it will yield libceres.a
4. grab the libceres.a
5. we will use the libceres.a as a static library for our project

Now, in our project, let's use the libceres.a and this is my Android.mk that uses libceres.a

  LOCAL_PATH := $(call my-dir)

  LOCAL_ALLOW_UNDEFINED_SYMBOLS := true

  include $(CLEAR_VARS)

  LOCAL_MODULE    := ceres

  LOCAL_SRC_FILES := libceres.a

  include $(PREBUILT_STATIC_LIBRARY)

  include $(CLEAR_VARS)

  LOCAL_C_INCLUDES += eigen
  LOCAL_C_INCLUDES += ceres-solver-1.7.0/include

  LOCAL_C_INCLUDES += ceres-solver-1.7.0/internal/ceres/miniglog

  LOCAL_MODULE    := myproject

  LOCAL_SRC_FILES := jni_part.cpp  

  LOCAL_STATIC_LIBRARIES  =  ceres

  LOCAL_LDLIBS +=  -llog -ldl 

  include $(BUILD_SHARED_LIBRARY)

Don't forget to change the path of the ceres-solver and eigen includes path.

In my error project, I use the linear-solver type DENSE_SCHUR as shown in the bundle adjustment tutorial. I also tried to change using SPARSE_NORMAL_CHOLESKY that can run faster but less accurate because I guess our problem is not dealing with sparse data. This solver can help to produce the camera pose even in the extreme tilt.

This is the snapshot of my implementation of shape tracking in android using ceres-solver. The camera pose can be computed as shown by the red outline in the images. Tested using Nexus 7.