kvmanohar22 / 3DR

Collection of tools for various vision related problems

3DR

Collection of tools ranging from image alignment, panorama generation, point cloud generation, optical flow estimation to single view 3D reconstruction.

Tracking camera (without Bundle Adjustment). Video: (KITTI `sequence/00`)

Usage

• Note: By default, build type is Debug, to change to Release, edit in CMakeLists.txt file
• To get log printed to screen, export the environment variable GLOG_logtostderr as;

export GLOG_logtostderr=1

1. Build all the executables

git clone [e-mail]:kvmanohar22/3DR.git
mkdir build
cd build
cmake ..
make

2. SLAM

• First download videos from KITTI dataset link
• Set the path to the video sequence in scripts/slam.sh

cd scripts/
./slam.sh

TODO

• Image Alignment
• Image Warping
• Panorama stitching
• Visual Odometry
• Semi-Direct Visual Odometry (SVO)
• Tracking
• Ghost removal in big panoramas (refer [1])
• Optical Flow
• SfM (unordered images)
• SfM (video sequence (localisation))

To fix

• Cholesky Decomposition fails during BA
• Optimization is ridiculously slow (but works). Should make it faster
• Fliped y-axis in 3D viewer?
• Add only KeyFrames for graph optimization
• Reduce the number of points for graph optimization

Semi-Direct Visual Odometry

• opencv's calcOpticalFlowPyrLK doesn't seem to work that good on forward motion camera Compute inliers from vikit seems to be doing it wrong! Fix it!
• Correspondance point estimation is noisy (probably). Currently using optical flow to estimate corresponding points.

Requirements

• C++14
• Linux (tested only on Ubuntu 18.10)
• OpenCV (for image I/O)
• ceres-solver (for Bundle Adjustment)
• Pangolin (for 3D viewer)
• fast (for fast corner detection)
• GLOG (for logging)

References

1. M. Uyttendaele, A. Eden, and R. Szeliski. Eliminating ghosting and exposure artifacts in image mosaics. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition

LICENSE

All my code is MIT licensed