peterzs / satelliteplanesweep

This implements plane sweep stereo for the satellite-based MVS.

Introduction

This code is adapted from PlaneSweepLib in order to reconstruct with satellite images. It can also work on ground-level images.

Project page: https://kai-46.github.io/VisSat/

Compilation

The following libraries are required:

NVIDIA CUDA
Boost (system filesystem program_options)
OpenCV
Eigen3

The script 'install_opencv.sh' aims to assist installing opencv on your computer.

Compile the program via:

mkdir build && cd build && cmake .. && make

General Usage

SatellitePlaneSweep/build/bin/satellitePlaneSweep \
    --imageFolder {} \
    --imageList {} \
    --refImgId {} \
    --outputFolder {} \
    --matchingCost {} \
    --windowRadius {} \
    --nX {} --nY {} --nZ {} \
    --firstD {} --lastD {} --numPlanes {} \
    --filterCostVolume {} --guidedFilterRadius {} --guidedFilterEps {} \
    --saveCostVolume {} \
    --debug {} \
    --saveBest {} --filter {} --filterThres {} \
    --saveXYZMap {} \
    --savePointCloud {}

Brief explanation of the options

• imageFolder: directory that contains your reference and source images
• imageList: a .txt file with each line being {img_id} {img_name} {fx fy cx cy s qw qx qy qz tx ty tz}; the number of lines should be equal to the number of reference and source images; {img_id} is an integer identifier, while {img_name} is the file name for the image inside {imageFolder}; {fx fy cx cy s qw qx qy qz tx ty tz} are the camera intrinsics and extrinsics
• refImgId: specify which image in the {imageList} you would like to be the reference image; the other images inside {imageList} will automatically become the source images
• outputFolder: where to save the program's output
• matchingCost: "ncc" or "census"
• nX, nY, nZ: they jointly define the normal direction of the sweeping planes in scene coordinate frame
• firstD, lastD: constants of the first and last sweep plane, respectively
• numPlanes: number of sweeping planes to be used
• filterCostVolume: whether to filter each slice of the cost volume with a guided-filter
• guidedFilterRadius, guidedFilterEps: parameters of the guided-filter
• saveCostVolume: whether to save the cost volume
• debug: if enabled, the program will output visualizations of the cost volumes and the warped images
• saveBest, filter, filterThres: whether to save the best plane hypothesis; before saving, an optional step filtering out the best plane hypothesis that have a cost above {filterThres} can be performed.
• saveXYZMap: whether to save the XYZ map, for which each pixel is the 3D point (X,Y,Z) in scene coordinate frame
• savePointCloud: whether to save the point cloud

Interfacing with Python

scripts/binary_grid_io.py can be used to read the output files of the C++ program as numpy array.

License

This software uses the GPLv3 license.