Slam-MapGenerator is a complete system to reconstruct surface from visual odometry (provided by online sparse SLAM / VO algorithm) + raw images.
For GSOC, You may refer to the [project wiki](https://jderobot.org/Club-jianxiong) for more progress-oriented content. (More images and videos there)
This project contains 3 major parts for now:
- Loop Closing
- Global Bundle Adjustment (BA)
- Reconstructing 3D Map
All 3 parts has finished development and testing. The whole program is capable of handing surface reconstruction in simple environment.
More complex environment would require further development.
We would discuss and continue further development after August 22rd, 2018. Feature enhancement and suggestions are more than welcomed.
The detailed README about datasets is in the subfolder datasets.
All datasets and some results are available at Google Drive. If the link doesn't work, you may type it manually: https://drive.google.com/drive/folders/1Ta6QgiQ5ASHj42pyGt9OijQ7t_oUynFb
If you are using the monocular dataset from TUM, which is fish-eye camera, you need to get undistorted images. You may do this by using the code provided in the TUM dataset webpage.
slam-MapGen requires the following dependency:
- cmake
- Eigen3
- OpenCV (3.0 or upper)
- [PCL Library] (http://www.pointclouds.org)
- DBoW2
- DLib
- Pangolin
- Ceres Solver
To install cmake, eigen3:
sudo apt-get install cmake libeigen3-devFor other dependency, you may follow the official installation guide.
Note: It is recommended to compile OpenCV3 from source and install. Using the anaconda-version with OpenCV3 will raise errors sometimes.
mkdir build
cd build
# If you don't want to view the debug info
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j4- Get the dataset from my google drive, choose TUM-RGBD-freiburg3-far, short-cut
- Download corresponding images from TUM (Sequence 'freiburg3_structure_texture_far'). short-cut
./build/map_generator ./config/example.yamlThe final reconstructed surface will be visualized with GUI.
pose-graph optimization result will be saved to pose_optimized.yaml (if enabled). Global Bundle Adjustment result will be saved to BA_optimized.yaml (if enabled). You may visualized them by slam-viewer
The second parameters of the above command is the configure file, it's in the following format:
The configure file has 4 blocks, detailed explanation are following:
| Parameter Name | Explanation |
|---|---|
| img_dir | The image directory (undistorted images) |
| trajectory | The input trajectory file |
| enableLoopClosure | Set to 1 to enable loop closure, set to 0 otherwise |
| enableBA | Set to 1 to enable global bundle adjustment, set to 0 otherwise |
| enableSurfaceRecon | Set to 1 to enable surface reconstruction, set to 0 otherwise |
Note: ${img_dir} + filename in trajectory file (YAML) should be the fullpath of the image file (JPG/PNG/BMP).
This is the reason that img_dir is written as /home/freiburg3_far/rgb/../
| Parameter Name | Explanation |
|---|---|
| Vocabulary | The pre-trained Visual Vocabulary File (You may get this through the above dataset link) |
| loop_detection_threshold | The threshold for deciding whether two frames are similar enough to be a loop closing pair |
| Parameter Name | Explanation |
|---|---|
| reconMethod | Only PlaneRANSAC+GreedyTriangulation is a reliable method working on test set for now. |
| enableStitchingImage | Set 1 to enable stitching image to texture on visualization, this would discard plane unable to be visualized. A plane is considered valid only if all 3 vertexes of the triangle appear on same single image |
| Parameter Name | Explanation |
|---|---|
| minPreserveRatio | The minimum inlier ratio RANSAC is going to preserve |
| Parameter Name | Explanation |
|---|---|
| mu | the multiplier of the nearest neighbor distance to obtain the final search radius for each point |
| maximumNearestNeighbors | the maximum number of nearest neighbors to be searched for. |
| searchRadius | the sphere radius used for determining the k-nearest neighbors |
| Note: detailed explanation refers to PCL Library GreedyTriangulation method |
All programs are tested under Ubuntu16.04.
We are using Github Issues to manage bugs and provide answers to questions. Feel free to file bug report / ask questions there. New feature request and discussion are also more than welcomed.
This project was proposed and accepted as a student program for GSOC 2018 (Google Summer of Code), from May 14th to August 14th, 2018. Detailed wiki for this project as GSOC 2018 on this project is on the project wiki.
The project will not get updated in short term, as the original author is not available for this project at this moment. New feature request and discussion are still more than welcomed.