Saliency Optimization from Robust Background Detection COMPUTER VISION LAB. 여동훈
Main Steps (a) Detect background boundary superpixel (b) Compute Distance from background (c) Saliency Optimization from (b) (a) (b) (c)
Detect background boundary superpixel
Without segmentation – soft segmentation Two terms Area of a region R Length along the boundary Compute the boundary connectivity for each superpixel The boundary connectivity of a soft region containing the superpixel
Bondary Connectivity of a superpixel
Boundary Connectivity Values on images (b) Edges link adjacent superpixels only (c) Edges additionally connect all pairs of boundary superpixels
Experimental Results on Boundary connectivity Thresholding Boundary connectivity with a single threshold (2) The values of the boundary connectivity is stable over the images
Main Steps (a) Detect background boundary superpixel (b) Compute Distance from background (c) Saliency Optimization from (b) (a) (b) (c)
Background Weighted Contrast Contrast Saliency Background Weighted Contrast
Comparison on two algorithms Contrast Saliency Background weighted Contrast
Main Steps (a) Detect background boundary superpixel (b) Compute Distance from background (c) Saliency Optimization from (b) (a) (b) (c)
Saliency Optimization
Measurement Precision – Recall curve Mean absolute error (MAE) Average per-pixel difference gt ours AE
Experiments BndCon Ctr wCtr wCtr*
MSRA-hard dataset 657 images where salient object touches image boundary
Optimization on Several Algorithms