Improving context-sensitive similarity via smooth neighborhood for object retrieval

Highlights：

• Aiming at improving context-sensitive similarity, we propose a novel algorithm for neighborhood structure mining, which satisfies the manifold assumption.

• The proposed Smooth Neighbourhood (SN) enables the neighbor selection to vary smoothly along the data manifold, thus the local similarity can be sufficiently reflected in the selection of neighbors.

• SN is suitable to deal with more than one affinity graph. It learns the shared neighborhood structure and the importance of multiple affinity graphs in a unified framework. Therefore, neighborhood aggregation and weight learning can be done simultaneously.

• Instead of using some heuristic rules that stem from empirical observations (e.g., Mutual kNN), we give a formal formulation to SN and derive an iterative solution to the optimization problem with proven convergence.