Continuous Depth Map Reconstruction From Light Fields

This project is for the research experiments for this paper [pdf]

Continuous depth map reconstruction from light fields
J. Li, M. Lu, and Z.N. Li,
IEEE Trans. on Image Processing, 24(11), 2015, 3257-3265.

The code base contains two executable parts. Part I is for initial depth estimation, which is described in section II and III in the paper. The part II is to optimize the initial depth map described in section IV in the paper.

Part I: Initial estimation

How to compile:

• Required libraries: OpenCV2 HDF VIGRA
  It might have some compatibility issues with OpenCV3. Sorry...
  If you are on mac:

  brew install homebrew/science/opencv
  brew install homebrew/science/vigra
  

• Compile:

  cd <root directory of this project>/initialEstimation
  OPENCV=<your opencv library directory> HDF=<your hdf library directory> VIGRA=<your vigra library directory> make
  

How to use:

• Input data:

  • Download data from the light field dataset provided by

    Datasets and Benchmarks for Densely Sampled 4D Light Fields
    S. Wanner, S. Meister, B. Goldluecke
    In Vision, Modelling and Visualization (VMV), 2013.

    At the writing time, they are updating the site and the data might not be available
    http://www.informatik.uni-konstanz.de/cvia/resources/.
    Old URL:
    http://hci.iwr.uni-heidelberg.de/HCI/Research/LightField/lf_benchmark.php

  • Put the h5 files under ./data/input/

• Run:

  cd <root directory of this project>/initialEstimation
  ./initial_estimation <data name>
  

• Output:
  Find the output data under ./data/initial_results/

Part II: Optimization

How to compile:

• Required libraries: OpenCV2 HDF Eigen

• Compile:

  cd <root directory of this project>/optimization
  OPENCV=<your opencv library directory> HDF=<your hdf library directory> EIGEN=<your eigen library directory> make
  

How to use:

• Input data:
  Make sure the output data from the first part is ready under ./data/initial_results/

• Run:

  cd <root directory of this project>/optimization
  ./optimize <data name> <use revised initial results> <optimization method> [<segment>]
  

  • data name: without extension. eg, buddha.
  • use revised initial results: 0 or 1
    Use revised initial result or original result from structure tensor
  • optimization method:
    NoOpt = 0: no optimization
    MRFOpt = 1: use Markov Random Field
    FastMatting = 2: use fasting matting to solve the linear equation
    ClassicOpt = 3: use Conjugate Gradient to solve the linear equation
    MixOpt = 4: use mixed method to solve the linear equation with segmentation
  • segment: 0 or 1
    This is only applicable for FastMatting, ClassicOpt and MixOpt

• Output:
  Find the output data under ./data/output/

Citation

• The project runs on the dataset provided by

Datasets and Benchmarks for Densely Sampled 4D Light Fields
S. Wanner, S. Meister, B. Goldluecke
In Vision, Modelling and Visualization (VMV), 2013.

• Mean Shift Analysis Library is included, which is based on these papers

D. Comaniciu, P. Meer: Mean Shift: A robust approach toward feature space analysis.
C. Christoudias, B. Georgescu, P. Meer: Synergism in low level vision.

• MRF energy minimization software is included, which was published accompanying this paper

A Comparative Study of Energy Minimization Methods for Markov Random Fields.
R. Szeliski, R. Zabih, D. Scharstein, O. Veksler, V. Kolmogorov, A. Agarwala, M. Tappen, and C. Rother.
In Ninth European Conference on Computer Vision (ECCV 2006), volume 2, pages 16-29, Graz, Austria, May 2006.