2017-CVPR-Instance-Level Salient Object Segmentation

Reference:

G. Li, Y. Xie, L. Lin and Y. Yu, Instance-Level Salient Object Segmentation, CVPR 2017. [project]

Overall Framework:

1. A mulit-scale model is proposed to compute the saliency map and contour map.
2. The MCG algorithms is used to find proposals.
3. The CRF is used to get the final segmentation of each instance.

MCG algoritms:

• 800 salient object proposals for any given image
• Discard those proposals with fewer than 80% salient pixels
• Apply a MAP-based subset optimization method proposed in [51] to produce a compact set of object proposals. (The final number of instances)

CRF optimization:

1. Suppose the number of salient instances is K, and the background is treated as K+1 class.
2. Define a probability map with K + 1 channels
   • If a salient pixel is covered by a single detected salient instance, the probability of the pixel having the label associated with that salient instance is 1.
   • If a salient pixel is not covered by any detected salient instance, the probability of the pixel having any label is 1/K .
   • If a salient pixel is covered by m overlapping salient instances, the probability of the pixel having a label associated with one of the m salient instances is 1/m .
   • If a background pixel is covered by m overlapping salient instances, the probability of the pixel having a label associated with one of the m salient instances is 1/(m+1) , and the probability of the pixel having the background label is also 1/(m+1).
3. A fully connected CRF is used for final optimization