Thesis Evaluation

Code used to evaluate the refinement methods investigated in our thesis work: Normalized Convolution Network and Dataset Generation for Refining Stereo Disparity Maps.

Link: http://liu.diva-portal.org/smash/get/diva2:1333176/FULLTEXT01.pdf

Info

3 methods for refining disparity maps were investigated:

1. nconv: Eldesokey et al. (2018) https://github.com/abdo-eldesokey/nconv
2. sdr: Yan et al. (2019) https://github.com/danielcranston/SDR
3. inpaint: Telea (2004) https://docs.opencv.org/3.3.1/df/d3d/tutorial_py_inpainting.html

These methods were evaluated on 2 datasets:

• middv3: Middlebury V3 training images: Scharstein et al. (2014) http://vision.middlebury.edu/stereo/data/
• liu: Our own dataset created as part of this thesis.

Excecuting the code:

1. Download the datasets by following the instructions found in the data/ folders README.md
2. To start the evaluation of a certain method on a dataset, execute python evaluate.py -mode MODE -set SET
   • replace MODE with 'ncconv', 'sdr' or 'inpaint'
   • replace SET with 'middv3' or 'liu'

Optional parameters to evaluate.py:

-epethresh : sets the threshold of the end-point-error map (in pixels)
             displayed when plotting is turned on. Defaults to 20.
-plot      : sets the plotting mode.  
				0=no plotting
				1=plots appear for each evaluation item
				2=plots appear for each evaluation item, and the figures are saved to disc

Notes

• Evaluation data from NConv is unfortunately only available for Middlebury V3.
• This code was made for my own convenience and does not cater to specific screen resolutions and matplotlib backends