Dense Semantic Correspondence

方法名称

度量相似性的方法

相似性的层面

针对问题

Canonical correspondence

(dis-)similarity metrics

raw pixel intensities

the same underlying scene,brightness   constancy assumption

Improved canonical correspondence

Robust (dis-)similarity metrics

raw pixel intensities

the same underlying scene,non-rigid   deformations、occlu-

sions、non-global intensity、constrast and colorimet-

ric changes

semantic correspondence

(SIFT Flow)

L₁ metric

sparse keypoint

matching,dense SIFT space(sacri
  ces some localization accuracy)

across scenes

Ours

Per pixel classifiers

raw pixel intensities(Dense)

across scenes