At the moment, we only support Polygon based geometry, which has some limitations:
The size of the polygon doesn't correspond to the pixel count of the object for two reasons: (a) the polygon itself is basically approximated from the LabelRegion and (b) the contour itself is not clearly defined, i.e. usually two sides (mainly right and down pixels) are ignored (as we find a contour on the pixel boundaries themselves).
We completely ignore holes in the LabelRegions.
Therefore we must offer a pixel-based feature set and cleary describe the limitations of the Polygon based feature-set (i.e. holes are ignored).
The pixel-based FeatureSet potentially only comprises features like: Number of Pixels, Hole count, Ratio of foreground/background inside the object, Average size of holes... what ever we can think of. Some of these features can be implemented in ops again.
dietzc
commented
8 years ago
At the moment, we only support
Polygonbased geometry, which has some limitations:LabelRegionand (b) the contour itself is not clearly defined, i.e. usually two sides (mainly right and down pixels) are ignored (as we find a contour on the pixel boundaries themselves).LabelRegions.Therefore we must offer a pixel-based feature set and cleary describe the limitations of the
Polygonbased feature-set (i.e. holes are ignored).The pixel-based
FeatureSetpotentially only comprises features like:Number of Pixels,Hole count,Ratio of foreground/background inside the object,Average size of holes... what ever we can think of. Some of these features can be implemented in ops again.