Skip to content

Latest commit

 

History

History
52 lines (41 loc) · 6.36 KB

new_features.md

File metadata and controls

52 lines (41 loc) · 6.36 KB
Raw

New Features

This is a list of the new features provided by PLVS:

  • Line segment detection, matching, triangulation and tracking with both pinhole and fisheye cameras (NEW experimental).
    • This capability can be enabled via the option Line.on in the yaml settings.
    • Removed some bugs, customized and optimized the adopted line_descriptor module from OpenCV library.
  • Dense reconstruction with different volumetric mapping methods: voxelgrid, octree_point, octomap, fastfusion, chisel, voxblox.
    • It can be enabled by using the option PointCloudMapping.on in the yaml settings and selecting your preferred method PointCloudMapping.type (see the comments in the yaml files).
    • With stereo cameras you can select the algorithm used to computed depth maps from stereo by setting StereoDense.type (GPU-accelerated libelas, libsgm, and other experimental OpenCV variants).
    • With fisheye cameras you can set StereoDense.needRectification: 1: This implies rectified stereo images are computed on-the-fly only in the point cloud mapping thread (not in the sparse mapping thread) along with depth maps (NEW experimental). Otherwise, when Camera.type: "KannalaBrandt8", you can alternatively set Camera.needRectification: 1 and rectify the input stereo images for both the sparse mapping thread and dense mapping thread (to this aim, use the new examples in the folder Examples).
  • Incremental segmentation with RGBD sensors and octree-based dense map.
    • It can be enabled by using the option Segmentation.on in the yaml settings of the RGBD cameras (only working when octree_point is selected as volumetric mapping method).
  • Augmented reality with overlay of tracked features, built meshes and loaded 3D models.
    • This viz can be enabled by using the button AR Camera in the viewer GUI.
    • A shader allows to viz points and lines with fisheye cameras.
  • Generated sparse and dense maps can be saved and reloaded.
    • You can save the generated sparse and dense maps anytime by using the GUI: first press the button Pause and then press the button Save. As a consequence, maps will be saved in the Scripts folder. In particular, (1) a sparse map will be always saved, (2) a dense map will be saved in the form of a ply (or of another custom format) only in the case you have set PointCloudMapping.on: 1. You can browse the produced ply point clouds by using CloudCompare.
    • Use the SparseMapping options (as showed in this TUM configuration file) in order to reload the sparse map. In particular, be sure to properly specify the SparseMapping.filename and then set SparseMapping.reuseMap: 1.
    • As for reloading the dense map, set PointCloudMapping.loadMap: 1 and configure PointCloudMapping.loadFilename.
  • Extraction of ORB keypoints via CUDA.
    • This capability can be optionally activated by using the option USE_CUDA in config.sh
  • Different methods can be used with calibrated stereo cameras for estimating depth maps: libelas, libsgm, opencv (these methods may need more fine tuning).
    • Use the option StereoDense.type to select your preferred method in the yaml settings for stereo cameras. This will work with your stereo datasets when PointCloudMapping.on is set to 1.
  • Some parts of the original ORBSLAM code were improved and optimized.
  • A new version of g2o is supported (tags/20230223_git). This can be enabled by setting the option WITH_G2O_NEW to ON in the main CMakeLists.txt of PLVS. Note that the new version of g2o will be automatically installed for you by the main build script (build.sh → build_thirdparty.sh → install_local_g2o_new.sh).
  • Smart pointers to manage points and lines (WIP for keyframes) are optionally available. See the file Pointers.h.
  • MapOjbect: Experimental representation for planar objects (WIP for PLVS II).
  • C++17 support. This can be configured at global level in config.sh by setting the variable CPP_STANDARD_VERSION.
  • Many convenient scripts are provided for launching apps, benchmarking and monitoring the system. See the Scripts and the Benchmarking folders.
  • Rerun visualizations are now supported (NEW).

Note: PLVS is an active project. The main README is under construction and will be updated soon with further information and details. Likewise, code improvements are coming soon.

You can find further details and videos on this page and in the following document:
PLVS: A SLAM System with Points, Lines, Volumetric Mapping, and 3D Incremental Segmentation 
Luigi Freda

Limitations

At present, we have some limitations with some specific sensor configurations.

  • Monocular sensors

    • Line features and volumetric reconstruction are not supported with monocular sensors.

  • Incremental segmentation: Incremental segmentation is only supported with RGBD sensors and octree-based dense map (PointCloudMapping.type: "octree_point"). Too fast movements can break the smoothness of the reconstructed surfaces and the resulting geometric segmentation may appear suboptimal.

Important notes

  • With pinhole stereo cameras (Camera.type: "Pinhole"), rectification is automatically applied when using the new examples in the folder Examples.
  • With fisheye cameras, Camera.type: "KannalaBrandt8", you can set Camera.needRectification: 1 and rectify the input stereo images for both the sparse mapping thread and dense mapping thread (to this aim, use the new examples in the folder Examples). Otherwise, when targeting a volumetric reconstruction, you can set StereoDense.needRectification: 1: This implies rectified stereo images are computed on-the-fly only in the point cloud mapping thread (not in the sparse mapping thread) along with depth maps.