Non-photorealistic video rendering

In this application, we take a painting, create an anisotropically blurred version of it, and then create animations of both paintings (blurred and non-blurred) rotating. The blurred painting sequence is treated as the training input, and the original painting sequence is treated as the training output. We can then take new sequences and produce videos that resemble both the painting sequences and the original video. Our method results in videos that are temporally coherent with stable backgrounds. Using nearest neighbor matching for this application results in videos that are more temporally incoherent, especially in static regions.

Can sequence, Da Vinci style

Training input Training output Test sequence Nearest neighbor output Our algorithm's output

Can sequence, Manet style

Training input Training output Test sequence Nearest neighbor output Our algorithm's output

Tennis sequence, Da Vinci style

Training input Training output Test sequence Nearest neighbor output Our algorithm's output

Tennis sequence, Manet style

Training input Training output Test sequence Nearest neighbor output Our algorithm's output