This project is a implementation of the rigid part of http://www.stanford.edu/~qianyizh/projects/elasticreconstruction.html, can be used for global optimization of point clouds registration.

The origin implementation can be found here: https://github.com/qianyizh/ElasticReconstruction

The paper: Elastic Fragments for Dense Scene Reconstruction

• PCL 1.6.x (http://www.pointclouds.org/)
• Boost 1.5.x (http://www.boost.org/)
• SuiteSparse (http://faculty.cse.tamu.edu/davis/suitesparse.html)
• METIS 4.x.x (http://glaros.dtc.umn.edu/gkhome/fsroot/sw/metis/OLD)
• Eigne 3 (http://eigen.tuxfamily.org/index.php?title=Main_Page)

This project runs on Windows 64bit System, all the lib above need to be vs2010-64bit.

MIT

1. Dump Depth Map

   In this step, we convert the *.vmap and *.nmap file to *.pcd, with filtering and clipping. Their's a bug in PCL 1.6.*, DON'T dump to PLY format.

    set BASE_DIR=.
    set POS_FILE=szs_pos.txt
    KALOFution.exe -run dumpmap -pos-file %BASE_DIR%\%POS_FILE% --dump-dir %BASE_DIR%\cloud-filtered --filter --data-dir %BASE_DIR%\dump --min-clip 0.5 --max-clip 2.4 --step 45
   

   • --pos-file: if provided, the output cloud will be translated according to camera poses. Usually used for debugging.
   • --dump-dir: where to store the cloud files. if the DIR doesn't exists, it will be created automatically.
   • --filter: if provided, the output cloud will be filtered using outlier removal.
   • --data-dir: where to load *.vmap and *.nmap.
   • --min-clip --max-clip: limit the points Z-axis to min and max.
   • --step: the dump step

2. Build Correspondence

   In this step, we find out all possible cloud pairs which have overlap to each other, and save their correnpondence point indexes. If cloud_x.pcd and cloud_y.pcd have correnspondence points, the result will be stored in file corres_x_y.txt

    set BASE_DIR=.
    set POS_FILE=szs_pos.txt
    KALOFution.exe -run buildcorres --pos-file %BASE_DIR%\%POS_FILE% --data-dir %BASE_DIR%\cloud-filtered --better-corres 1 --need-align 1 --step 45 --cloud-num 149 --save-to %BASE_DIR%\corres-better --dist-thres 0.05 --angle-thres 15 --valid-pair-dist-thres 0.03 --align-sampling-limit 30000 --min-corres-num 8000 --align-translation-thres 0.3
   

   • --pos-file: the camera poses.
   • --data-dir: where to load *.pcd.
   • --better-corres: use better correspondence measurement (slower) or not (faster).
   • --need-align: use ICP point-to-plane alignment before finding correspondent points (much slower) or not (faster).
   • --cloud-num: how many clouds to load.
   • --step: the loading step, the same as Dump Depth Map.
   • --save-to: where to store corres_x_y.txt.
   • --dist-thres: the maximum correspondent points distance for alignment.
   • --angle-thres: the maximum correspondent points norm angle (in degree) for alignment.
   • --valid-pair-dist-thres: the maximum distance for a valid correspondence pair.
   • --align-sampling-limit: down sample number for alignment.
   • --min-corres-num: the minimum points pairs for a valid cloud pair.
   • --align-translation-thres: the minimum translation distance for a valid alignment.

3. Global Optimization

   In the step, we fill a sparse matrix according to the conrespondence relations, and solve the equation to minimize the global alignment error.

    set BASE_DIR=.
    set POS_FILE=szs_pos.txt
    KALOFution.exe -run optimize --pos-file %BASE_DIR%\%POS_FILE% --data-dir %BASE_DIR%\cloud-filtered --step 45 --cloud-num 149 --corres-dir %BASE_DIR%\corres-better --save-to %BASE_DIR%\optimized-better
   

   • --pos-file: the camera poses.
   • --data-dir: where to load *.pcd.
   • --corres-dir: where to load cloud_x_y.txt.
   • --save-to: where to store the optimized cloud files (in PLY format).