New light field depth estimation benchmark

We recently published a new light field benchmark at ACCV 2016 in a collaboration with the HCI. The main  features are

  • a new synthetic dataset with 24 carefully designed scenes, which overcomes technical shortcomings of previous datasets.
  • novel error measures and evaluation modalities enabling comprehensive and detailed characterizations of algorithm results.
  • an initial performance analysis of four state-of-the-art light field algorithms and one multi-view stereo algorithm.
  • a benchmarking website and an evaluation toolkit to provide researchers with the necessary tools to facilitate algorithm evaluation.

The benchmarking website is hosted at lightfield-analysis.net. Please refer to our ACCV paper in case you use any of the datasets.

 

HCI light field database

The HCI light field database recently went offline after their web page was restructured. As we co-authored the paper and there is still lots of demand for them, we temporarily host the data sets on our own servers.

The following script points to modified download locations:

Modified download script for HCI light field database

Please refer to our VMV paper details, and cite it in case you use any of the datasets.

 

Code

We develop the open source library cocolib, a C++/CUDA toolkit for variational image analysis and continuous global optimization. The focus lies on the minimization of functionals of the type

\( E(u) = J(u) + F(u) \)

where \( u \) is an image (i.e. vector-valued function on a grid), \( J \) is a convex regularizer, and \( F \) a (usually) convex data term. The library comes with a command line tool and Matlab interface, as well as example code for most of the available algorithms.

A number of generic and more specialized solvers are implemented, focusing on linear inverse problems and multi-label segmentation with various models for regularization and data terms. Also included is a suite for variational light field analysis, which ties into the HCI light field benchmark set and givens reference implementations for a number of our recently published algorithms.

Please see the project home page for more information.

Note that the project is orphaned since 2015 - as Matlab has a very decent GPU support, we now usually use a combination of Matlab and CUDA C++ plugins for development.