Skip to content

Latest commit

 

History

History
105 lines (81 loc) · 3.83 KB

matching_cost.rst

File metadata and controls

105 lines (81 loc) · 3.83 KB
Raw

Matching cost computation

Theoretical basics

The first step is to compute the cost volume containing the similarity coefficients. Different measures of similarity are available in Pandora :

  • SAD (Sum of Absolute Differences)
  • SSD (Sum of Squared Differences)
  • Census [Zabih1994]
  • ZNCC (Zero mean Normalized Cross Correlation)

It is possible to oversample the cost volume by a factor of 2 or 4 ( with the subpix parameter ) compared to to the input images. It can be useful for :ref:`disparity_refinement`

Note

The cost volume disparity dimension is sampled at the input images rate by default. Thus disparities tested are integers. However, to prevent from aliasing effects when refining the disparity map, one can use the subpix parameter. This will add subpixel disparities into the cost volume by oversampling the disparity dimension by an even factor.

Pandora can take into account a mask and a nodata value for each image. The masks and nodata are used during the matching cost computation :

  • Nodata pixel management: if the left window contains nodata, the center pixel of the window is invalidated. Therefore,the disparity range is invalidated : cost(x, y, \forall d) = nan. If the right window contains nodata, the center pixel is invalidated. As a result, the pixels of the left image such as I_{L}(x, y) = I_{R}(x + d, y), are invalidated cost(x, y, d) = nan
  • Management of hidden pixels: if the center pixel of the left window is hidden, the disparity range is invalidated : cost(x, y, \forall d) = nan. If the pixel in the center of the right window is hidden, the pixels of the left image such as I_{L}(x, y) = I_{R}(x + d, y) are invalidated cost(x, y, d) = nan
[Zabih1994]Zabih, R., & Woodfill, J. (1994, May). Non-parametric local transforms for computing visual correspondence. In European conference on computer vision (pp. 151-158). Springer, Berlin, Heidelberg.

Configuration and parameters

Name Description Type Default value Available value Required
matching_cost_method Similarity measure string   ssd sad census zncc or mc_cnn if plugin_mccnn is installed Yes
window_size Window size for similarity measure int 5 Must be >0 and For "census" : {3 5} No
subpix Cost volume upsampling factor int 1 {1 2 4} No
band Band value on which to calculate the correlation str None band in left and right image's metadata No
step Computing step for pandora2d int 1 1 No

Note

  • NB: Parameter band_im must be present in the left and right image's metadata.
  • If the input images are multiband, the band's names must be present on the image metadata. To see how to add band's names on the image's metadata, please see :ref:`faq`.

Example for mono band images

{
    "input" :
    {
        // ...
    },
    "pipeline" :
    {
        // ...
        "matching_cost":
        {
            "matching_cost_method": "ssd",
            "window_size": 7,
            "subpix": 4
        }
        // ...
    }
}

Example for multiband images

{
    "input" :
    {
        // ...
    },
    "pipeline" :
    {
        // ...
        "matching_cost":
        {
            "matching_cost_method": "ssd",
            "band":"r",
            "window_size": 7,
            "subpix": 4
        }
        // ...
    }
}