Projects‎ > ‎

(Co-)Salient Object Detection by Rank-Sparsity Decomposition

No segmentation...but this work can still detect "near-segment" salient objects. The success is mainly based on
  • Modeling non-salient parts by low-rank approximation; and
  • Using "sparse support" for saliency-aware decomposition to allow spatial coherence.
The (spatial) coherence is embedded in a graph-based sparse matrix and used for quadratic programming (QP) to achieve "near segment" optima. 

Following this technique, we generalize our method to co-salient object detection by constructing a joint-graph and thereby exploit the (feature-like) coherence between images.