Temporally Coherent Completion of Dynamic Video

摘要

We present an automatic video completion algorithm that synthesizesrnmissing regions in videos in a temporally coherent fashion.rnOur algorithm can handle dynamic scenes captured using a movingrncamera. State-of-the-art approaches have difficulties handlingrnsuch videos because viewpoint changes cause image-space motionrnvectors in the missing and known regions to be inconsistent. Wernaddress this problem by jointly estimating optical flow and colorrnin the missing regions. Using pixel-wise forward/backward flowrnfields enables us to synthesize temporally coherent colors. We formulaternthe problem as a non-parametric patch-based optimization.rnWe demonstrate our technique on numerous challenging videos.