Coupling Dense Saliency with Mesh Constraints Across Scales

Determine how dense saliency should be coupled to mesh constraints across scales in learning-based image retargeting so that semantic visual attention and sparse geometric deformation remain effectively aligned.

Background

Learning-based image-retargeting methods combine semantic representations with explicit geometric deformation, while recent approaches also use self-supervised mesh warping and separate deformation fields for salient and non-salient regions. However, dense saliency is defined at the pixel or object-boundary level, whereas mesh warping is controlled by sparse vertices and cells.

The paper identifies the unresolved issue as determining an effective cross-scale coupling between these dense visual-attention priors and sparse mesh constraints. AttenIR addresses this issue for still images by projecting mesh-induced losses into the pixel domain, but the broader problem of how such coupling should operate across scales remains explicitly open in the related-work discussion.