Using Photographs to Enhance Videos of a Static Scene Pravin Bhat 1, C. Lawrence Zitnick 2, Noah Snavely 1, Aseem Agarwala 3, Maneesh Agrawala 4, Michael Cohen 1,2, Brian Curless 1, Sing Bing Kang 2 EGSR 2007 University of Washington 1, Microsoft Research Redmond 2 University of California 3, Adobe Systems 4
An overview of Spacetime Fusion
Motivation Low quality video Input Video
Motivation Low quality video Reconstructed video Input VideoReconstructed Video
Motivation Low quality video Reconstructed video –Reconstructed from photos –Good spatial reconstruction –Bad temporal reconstruction Input VideoReconstructed Video
Motivation Spacetime Fusion result Input VideoSpacetime Fusion Result
Motivation Spacetime Fusion result –Spatial properties of reconstruction –Temporal properties of input video Input VideoSpacetime Fusion Result
Define a 3D gradient field Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Video frame: t Video frame: t - 1 Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Video frame: t Video frame: t - 1 GtGt G t (x, y, t) = V(x, y, t) - V(x, y, t - 1) Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Video frame: t Video frame: t - 1 GtGt G t (x, y, t) = V(x, y, t) - V(x - u, y - v, t - 1) Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Increases compatibility between temporal gradients and spatial gradients Spacetime Fusion
Define a 3D gradient field –Spatial gradients from reconstruction –Temporal gradients from input video –Key Idea Temporal gradients defined between motion compensated temporal neighbors Increases compatibility between temporal gradients and spatial gradients Integrate the 3D gradient field Spacetime Fusion
Integrating the gradient field Solve linear system: Av = b
Integrating the gradient field Solve linear system: Av = b Constraints: v x, y, t – v x-1, y, t = G x (x, y, t) v x, y, t – v x, y-1, t = G y (x, y, t) v x, y, t – v x-u, y-v, t = G t (x, y, t) Spacetime Fusion
Applications
Enhanced Exposure
Input Video Edit Propagation
User Edits
Edit Propagation User Edits
Edit Propagation User Edits
Edit Propagation User Edits
Edit Propagation User Edits
Edit Propagation
Edited Video Edit Propagation
Super-Resolution
Conclusion Spacetime fusion
Conclusion Spacetime fusion –Combines spatial and temporal gradients from two different sources
Conclusion Spacetime fusion –Combines spatial and temporal gradients from two different sources –Requires motion vectors for temporal source stereo (static scenes) flow (dynamic scenes)
Conclusion Spacetime fusion –Combines spatial and temporal gradients from two different sources –Requires motion vectors for temporal source stereo (static scenes) flow (dynamic scenes) –Major applications Enforcing temporal coherence Transferring lighting information