Semi Global Boundary Detection (SGBD) breaks the image into scan lines in multiple orientations, segments each one independently, and combines the results into a final probabilistic 2D boundary map. The reason we break the image into scan lines is that they are a mid-level image representation that captures more information than a local patch. Luckily, scan lines are 1D signals that can be optimally segmented using dynamic programming, and we make the assumption that the border pixels between segments are boundary pixels in the image. This leads to a simple and efficient algorithm for boundary detection. The entire algorithm requires little parameter tuning, works well across different data sets and modalities, and produces sharp boundaries. We report results on several benchmarks that include both color and depth.
- Edge/Boundary detection
- Semi global matching