sky-github / stereo_sparse_depth_fusion

Repository for "Real Time Dense Depth Estimation by Fusing Stereo with Sparse Depth Measurements" | OpenCV, C++

Stereo and Sparse Depth Fusion

This is the corresponding code repository for our paper Real Time Dense Depth Estimation by Fusing Stereo with Sparse Depth Measurements

Reference

If you use this work, please cite our paper:

@article{shivakumar2018real,
  title={Real Time Dense Depth Estimation by Fusing Stereo with Sparse Depth Measurements},
  author={Shivakumar, Shreyas S and Mohta, Kartik and Pfrommer, Bernd and Kumar, Vijay and Taylor, Camillo J},
  journal={arXiv preprint arXiv:1809.07677},
  year={2018}
}

In this repository, we also use the Semi Global Matching algorithm implementation by Koichiro Yamaguchi, therefore if you use this repository, please also cite the original paper by Koichiro et al. and follow their original license agreements.

@inproceedings{yamaguchi2014efficient,
  title={Efficient joint segmentation, occlusion labeling, stereo and flow estimation},
  author={Yamaguchi, Koichiro and McAllester, David and Urtasun, Raquel},
  booktitle={European Conference on Computer Vision},
  pages={756--771},
  year={2014},
  organization={Springer}
}

Dependencies

1. OpenCV (Tested on 3.X.X)
2. libpng++ (Tested on 0.2.5-1)

sudo apt-get install libpng++-dev

3. OpenMP (optional)
4. CMake

Testing Code

Navigate to the main code repository

mkdir build
cd build
cmake ..
make -j

You can run our KITTI example as follows

./stereo_depth_fusion

You should now see the following images in the results folder:

• sgm_default.png - Output of SGM algorithm
• sparse_mask.png - Mask of ground truth disparities used
• fuse_naive.png - Output of our naive fusion implementation
• fuse_neighborhoodsupport.png - Neighborhood Support Method
• fuse_diffusionbased.png - Diffusion Based Method

Understanding The Code

The entire Semi Global Matching implementation is from Koichiro Yamaguchi and lives in the SGMStereo.cpp and SGMStereo.h files. The adjustments to the cost volumes are made in SGMStereo.cpp in the following functions:

• updateCostVolume_NRF() - Naive Fusion
• updateCostVolume_NS() - Neighborhood Support
• updateCostVolume_DB() - Diffusion Based Method

These methods are called by a helper script stereo_depth_fusion.cpp and is tested on a pair of randomly chosen KITTI 2015 images.

Results

KITTI 2015

From top to bottom:

• Input color image (left)
• Semi Global Matching (original)
• Neighborhood Support Method (ours)
• Diffused Based (ours)
• Anisotropic Diffusion
• Ground Truth (LiDAR)

Images:

Middlebury

From left to right:

• Input color image (left)
• Semi Global Matching (original)
• Neighborhood Support Method (ours)
• Diffusion Based (ours)
• Anisotropic Diffusion
• Ground Truth

Images:

PMD Monstar Dataset (internal)

From left to right:

• Input color image (left)
• Semi Global Matching (original)
• Neighborhood Support Method (ours)
• Diffusion Based (ours)
• Anisotropic Diffusion
• Mask of Time-of-Flight depth measurements

Images:

If you have any questions regarding this repository or our method feel free to raise an issue request in this repository or email me directly.

Note: this repository is part of a larger on-going research project, and therefore contains multiple variables that are unused (but are used in following research). I've tried to add ignores to most of them, but please excuse these if they occur. Feel free to make pull requests if you notice any bugs or wish to cleanup this code.