This code is an implementation of RGB-D plane detection and color-based plane refinement with MRF(graph-cut) optimization.
Example: input color and depth image of frame-000000 from BundleFusion dataset's copyroom (the depth image here is scaled by a factor 10 for rendering purpose):
Result planes without (left) and with (right) MRF-based optimization:
General pipeline:
@inproceedings{wang2018plane,
title={Plane-Based Optimization of Geometry and Texture for RGB-D Reconstruction of Indoor Scenes},
author={Wang, Chao and Guo, Xiaohu},
booktitle={2018 International Conference on 3D Vision (3DV)},
pages={533--541},
year={2018},
organization={IEEE}
}
and plane detection on a single RGB-D frame:
@inproceedings{feng2014fast,
title={Fast plane extraction in organized point clouds using agglomerative hierarchical clustering},
author={Feng, Chen and Taguchi, Yuichi and Kamat, Vineet R},
booktitle={Robotics and Automation (ICRA), 2014 IEEE International Conference on},
pages={6218--6225},
year={2014},
organization={IEEE}
}
- OpenCV 3
- Eigen 3
- MRF 2.2 (already included)
- Fast plane detection on a single depth image: PEAC (already included)
RGBDPlaneDetection <-o> color_image depth_image output_folder
-o (optional) is to run MRF-based optimization for plane refinement. For example:
RGBDPlaneDetection ../pic/frame-000000.color.jpg ../pic/frame-000000.depth.png ../pic
RGBDPlaneDetection -o ../pic/frame-000000.color.jpg ../pic/frame-000000.depth.png ../pic
Two scripts demo_win.sh and demo_linux_mac.sh are provided for reference to run the code on a RGB-D sequence with multiple images. Note to modify the corresponding paths.
-
Windows: use Visual Studio to open sln file and compile and build the code. It is tested successfully in Visual Studio 2010 and 2013, and should work on all other Visual Studio platforms. Note to change your OpenCV and Eigen 3 paths if needed.
-
Linux or Mac OS: run
build_linux_mac.shto build the code.
Without MRF optimization, the execution program by Visual Studio 2010 or 2013 runs at about 25 FPS (including data I/O) on RGBD images with resolution 640x480 in a PC with 16GB RAM and intel i7 processor. With MRF optimization, the same code runs much slower at about 7 seconds per frame on the same data.
Input: color image and depth image.
Output:
- Plane segmentation image in PNG;
- Plane label file in TXT: the label index of the plane which each pixel belongs to, starting from 0 to N - 1 where N is the number of planes. If a pixel is not on any plane, then its label value is N.
- Plane data file in TXT. Each line represents one plane with format like this:
#plane_index(starting from 0 to #planes - 1) number_of_points_on_the_plane plane_color_in_png_image(r,g,b between [0,255]) plane_normal(1x3) plane_center(1x3) sx sy sz sxx syy szz sxy syz sxz
Here (sx sy sz) are average of sum of all 3D points (x, y, z) on the plane, (sxx syy szz sxy syz sxz) are the average of sum of x*x, y*y, z*z, x*y, y*z, z*z of all 3D points on the plane, respectively.
- Without MRF optimization, this code is simply the re-use of PEAC code.
- Currently the code only works on BundleFusion or 3DLite depth data format (16UC1 image with depth value in millimeter). If you want to use other kinds of RGB-D data, you need to rewrite the part of reading color and depth images, reset the camera intrinsic parameters the scale factor for depth images in
plane_detection.h. - NOTE for MRF crash in Windows: Sometimes MRF 2.2 source code crashes in Visual Studio in Windows due to some kind of memory leak problem, but this code works well in Linux and Mac OS. If you meet this problem, just have a try in Linux or Mac OS, or remove the MRF optimization option
-o.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent