Computing importance of 2D contour parts by reconstructability

Part-based representation has become an essential representation for objects. Different parts of a shape have various importance to shape perception. Concerned by the contributions of local parts to the perception of global shapes, we introduce an importance measure of shape parts from the viewpoint of their “reconstructability”. Given one shape part as a local observation, we estimate the global object shapes through a proposed novel shape reconstruction method. The reconstruction quality is determined by part geometry variation and part uniqueness. Then a conditional entropy formulation is introduced to measure the part importance considering the reconstruction quality, part uniqueness and variation. Extensive experiments demonstrate the benefit of the proposed part importance to object detection applications. Moreover, psychological experiments are conducted to further support the consistency of our importance measure with human perception towards shape parts.