RohanChacko / 3d-reconstruction-from-accidental-motion

Implementation of 3D reconstruction from accidental motion, CVPR 2014

Home Page:https://ieeexplore.ieee.org/abstract/document/6909904

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

3D Reconstruction from accidental motion

Reconstructs a 3D scene by generating a dense depth map from a set of images capturing a scene.

The algorithm first uses KLT Tracking to track features across all images. We use Bundle Adjustment using the Ceres Solver and estimate the extrinsic camera parameters. Finally, we solve a smooth dense map using a DenseCRF formulation. Implemented as a part of the Computer Vision course at IIIT H. Overview

Building Ceres Solver

  • mkdir ceres-bin
  • cd ceres-bin
  • cmake ../ceres-solver-1.11.0/
  • make -j4 or
  • bash recompile.sh

Install dependencies and requirements

  • pip install -r requirements.txt

Running KLT tracking and Bundle Adjustment

  • cd src
  • python main.py

Running DenseCRF

  • python dense_crf.py --min_d 2 --max_d 4 --scale 2 --patch_rad 1 --iters 100 --p_std 3,3 --c_std 20,20,20 --wt 1 --max_p 0.15 --folder stone4 --nsamples 64 --show_wta

  • If plane sweeping is already computed for the scene, add the flag --pc <path-to-npz>

Outputs

  • cd outputs

Building DenseCRF package

  • cd src
  • chmod +x src/pydensecrf_setup.sh
  • ./pydensecrf_setup.sh

Note: Make sure Cython is installed before running the above commands.

Dataset Location

  • Download the dataset from here and store it in ./datasets

TODO

  • KLT
  • Filtering Good Points
  • Bundle Adjustment
  • CRF Energy Minimization
  • Integration

About

Implementation of 3D reconstruction from accidental motion, CVPR 2014

https://ieeexplore.ieee.org/abstract/document/6909904


Languages

Language:C++ 89.5%Language:Jupyter Notebook 5.4%Language:CMake 2.6%Language:Python 1.4%Language:C 0.8%Language:Makefile 0.2%Language:Shell 0.1%Language:HTML 0.0%