Minjae Lee (minjaele) 15-463 Computational Photography Final Project

Video texture

Approach
1. Finding good transitions
Compute L_2 distance D_(i, j) between all frames (assuming that similar frame makes good transitions)
Cost function: C_(i -> j) = D_(i + 1, j)
Probability for transition P_(i -> j) inversely related to cost: P_(i -> j) ~ exp(-C_(i -> j) / sig^2)
original media
2. Preserving dynamics + Some offset adjustment to make it look cool
Cost function: C_(i -> j) = sum_(k = -N)^(N - 1) (w_k * D_(i + k + 1, j + k))
(Use gaussian for weights?)
original media

Single-view Reconstruction

Approach
Matlab Part
1. Get vanishing point and coordinate values for each corner
2. Warp each sides to a texture map
3. Compute depth
4. Output stuff
C++ Part
1. Parse textures and map them to cube
2. BOOM (only on windows)

* Source code ripped from famous Nehe tutorial


Original Image:


Results:




Original Image:


Results: