Location Shift and Extent Trace
I found what I think has been done by other guys. This idea comes from Zhang Peng and Wang Run-Sheng of National University of Defense Technology, China. The process of salient region detection is devided by Location Shift and extent trace.
Location Shift: Using space-based attention model to select next salient location;
Extent Trace: In each salient location, use a set of continue large DoG templates to calculate the spatial center-surround contrast. The output of these masks generate a curve of local spatial neighboring contrasts of the region with different size. Intuitively, the scale with highest contrasts estimate the region size.
The idea is simple but reasonable. It avoids any region extraction algorithm on saliency map but provides a analytic method to estimate the salient region size that is lack in space-based visual attention model.
Location Shift: Using space-based attention model to select next salient location;
Extent Trace: In each salient location, use a set of continue large DoG templates to calculate the spatial center-surround contrast. The output of these masks generate a curve of local spatial neighboring contrasts of the region with different size. Intuitively, the scale with highest contrasts estimate the region size.
The idea is simple but reasonable. It avoids any region extraction algorithm on saliency map but provides a analytic method to estimate the salient region size that is lack in space-based visual attention model.
