Skip to content

Repository for the ICCV 2019 paper: View-consistent 4D Light Field Superpixel Segmentation, by Khan et al.


Notifications You must be signed in to change notification settings


Repository files navigation

View-Consistent 4D Light Field Superpixel Segmentation

Numair Khan1, Qian Zhang1, Lucas Kasser1,Henry Stone1, Min H. Kim2, James Tompkin1
1Brown, 2KAIST
ICCV 2019

Presentation Video (YouTube)


If you use this code in your work, please cite our paper:

  title={View-consistent 4D Lightfield Superpixel Segmentation},
  author={Numair Khan, Qian Zhang, Lucas Kasser, Henry Stone, Min H. Kim, James Tompkin},
  journal={International Conference on Computer Vision},

Running the MATLAB Code

Generating Superpixels

To generate a superpixel segmentation of a light field, use followed by the path to the light field file:

sudo ./ <path-to-light-field>

The light field is provided as a .mat file containing a 5D array. The dimensions of the 5D array should be ordered as (y, x, rgb, v, u) where "rgb" denotes the color channels.

       |  ________    ________
       | |    x   |  |    x   |
       | |        |  |        |
     v | | y      |  | y      | ....
       | |        |  |        |     
       | |________|  |________| 
       |           :
       |           :

Alternatively, a path to a directory of images may be provided to The directory should contain a file called config.txt with the dimensions of the light field on the first line in format y, x, v, u.

Make sure to set the camera movement direction for both u and v in parameters.m.

The segmentation results are output to a 4D MATLAB array in ./results/<time-stamp>/.

Comparing Your Own Results on the HCI dataset

It's easy to compare results of your own light field superpixel algorithm to our method on the old HCI dataset. Since the old HCI dataset provides ground truth segmentation labels and disparity maps, we can evaluate results both quantitatively and qualitatively.

To do this, first download and extract the dataset to a location on your file <path-to-hci-dataset>. Then run:

sudo ./ <path-to-your-results> <path-to-output-plots> <path-to-hci-dataset>

This will output qualitative and quantitative comparison figures at location <path-to-output-plots>. This takes some time to finish, so you may want to run it overnight.

<path-to-your-results> should contain your results stored as 4D arrays in separate .mat files, with the following considerations:

  • The filename of each .mat file should contain the name of the HCI light field it stores labels for (buddha, horses, papillon, or stillLife -- casing doesn't matter).

  • Provide separate files for each superpixel size.

  • Make sure your 4D arrays follow the dimensional order (y, x, v, u)

Comparing your own results on other datasets

For datasets other than the old HCI, we can only do qualitative comparison. See the function eval/plotComparisonQualitative.m for doing this from MATLAB.


  • Code fails with error Index exceeds the number of array elements: Make sure you are following the correct dimensional ordering; for light field images this should be (y, x, rgb, v, u) and for superpixel labels (y, x, v, u).
  • The segmentation output has very high error: Make sure you specify the direction in which the camera moves in u and v. This can be done by setting the boolean variables uCamMovingRight and vCamMovingRight in parameters.m. The camera movement direction determines the occlusion order of EPI lines, and is important for the bipartite matching stage of our algorithm.
  • The code has been run and tested in MATLAB 2019a. Older version of MATLAB may throw errors on some functions.


  • In Figure 8 of the main paper, the central row of the EPFL light fields show view(4, 4), rather than view(5, 5).
  • The manner in which labels are propagated has been updated in this codebase. Before, we began at the top-most view in the central column and moved down. In this code, we move out from the central view in a spiral. This leads to a slight improvement over the results reported in the ICCV published paper. The paper linked from this repository includes the new results which match the code repository.
  • The supplemental material linked from this repository has additional analysis of hyperparameter variation for all techniques across our analysis metrics.


Repository for the ICCV 2019 paper: View-consistent 4D Light Field Superpixel Segmentation, by Khan et al.







No releases published


No packages published