key point predictor: Predict K key points and corresponding heat maps
motion animation
coarse motion estimation: predict K affine transform between X -> S and X -> T. calculate the affine transform (S -> D)of each pixel by weighted sum(softmax heat maps + predicted affine transform)
dense motion estimation: use the affine transform per pixel to predict the dense optical flow and confidence map
image generation: use optical flow to warp the feature map and apply confidence map
standard: for D1, D2, ,,,, Dt estimate motion by S->D1, S->D2, ... S->Dt
relative: for D1, D2, ,,,, Dt estimate motion by S->D1, S+(D1->D2), ..., S+(D1->Dt)
Introduction
Method
estimate motion from S -> T, arbitrary X
key point predictor: Predict K key points and corresponding heat maps
motion animation
image generation: use optical flow to warp the feature map and apply confidence map
standard: for D1, D2, ,,,, Dt estimate motion by S->D1, S->D2, ... S->Dt
relative: for D1, D2, ,,,, Dt estimate motion by S->D1, S+(D1->D2), ..., S+(D1->Dt)
Highlight
Limitation
Comments