hitdshu / hitnlls

A nonlinear least square(NLLS) solver. Fomulate the NLLS as graph optimization.

hitnlls

A nonlinear least square(NLLS) solver. Fomulate the NLLS as graph optimization. For the name, HIT here refers to Harbin Institute of Technology to honor the six wonderful years spent there and NLLS is short for nonlinear least square.

Features

• This library is self contained and only requires a c++ 11 compiler. It's dependency free.
• A highly efficent matrix header-only library that uses template-meta programming and expression templates to reduce temporary variables.
• Support dense solvers, sparse solver and schur solver.
• Auto differentiation exploiting dual numbers are available both for vector variables and manifolds(SE3 etc.).
• A light weight geometry library, containing several lie groups, triangulizations, etc..
• A light weight admm solver for convex cone optimization problem.

Abstract

This solver uses Gauss-Newton type algorithms for NLLS problems. We define node to be variables waiting for optimization and factor to be observation over which jacobians and errors are evaluated. Without any dependence on other libraries, we choose to implement dense matrix and sparse matrix operations used in later computation. We've updated this new version of HITNLLS, which extensively uses template metaprogramming and CRTP to make the code more structured, more efficient (and more difficult to read). If you want to see old version of HITNLLS, check the other branch of the repo. Currently we're adding unit tests and examples.

Usage

install

$ git clone https://github.com/hitdshu/hitnlls.git
$ cd hitnlls && mkdir build && cd build
$ cmake ..
$ make
$ sudo make install
$ make test # optional

usage in cmake

find_package(hitnlls REQUIRED)
include_directories(
${hitnlls_INCLUDE_DIRS}
)
target_link_libraries(Test 
${hitnlls_LIBRARIES})

Implementation

admm

We developed a cone ADMM algorithm for convex cone optimization problem(LP/SOCP/SDP). Check files there if you feel interested.

camera

Several camera models are defined here. Our implementation of cameras are integrated naturally into our optimization framework.

common

This folder defines the register and timer, which are needed by other modules. Included also is a thread-safe logger.

geometry

Our customized implementation of several lie groups(SO2/SO3/SE2/SE3) can be utilized for representation and optimization of poses in 2d and 3d space. The jet class is for auto differentiation just as ceres. The lie groups defined here can be combined with jet to make auto differentiation in mainfolds much simpler.

matrix

We use CRTP and expression templates to enable lazy evaluation and avoid the generation of temporaray variabls. Our api of dense matrix operations resembles those of Eigen. We mainly implement cholesky/qr/inverse/evd/svd/lup operations for dense matrix.

ils

This folder contains optimizable problem which organizes nodes and factors and optimization algorithms. We implement GAUSS_NEWTON, LEVEN_MARQ and TRUST_REGION strategies and DENSE_QR, DENSE_CHOLESKY, SPARSE_CHOLESKY and DENSE_SCHUR algorithms in this folder.

type

Factors and vertices are defined here. We use variable tuples to avoid dynamic_cast in the evaluation of factors. The framework can support hand diff/auto diff/numeric diff now. Inherits Factor/FactorAutodiff/FactorNumeDiff according to your own purpose.