Texture Synthesis for Perspective Example

Before the final report, I gives a report about some progress after the presentation on November 20th. The first part is to optimize current algorithm’s time and space complexity. I am rewriting the KD-tree data structure used in patch match process. There are two main differences. First, in previous KD-tree implementation, dim-n vectors are stored in the KD-tree locally, which makes the space complexity of tree construction very high. In fact, it is impossible to set neighbor size more than 60 on my 1G RAM laptop. In my new implementation, for one patch’s dim-n feature vector, I only store pointer (or index) of the patch’s center pixel and the dim-n feature vector will be compute in-place based on the image data. Second, I apply the basic PCA dimension reduction operation on the high dimension feature vector, and reduces them from dim-n to dim-m, where m n. I find the dimension reduction will bring several advantages: 1) the time complexity for inplace feature computation is less; 2) the memory consume is much less; 3) the match becomes more robust: because many feature noise are filtered and the match will focus more on main features. I am also working on improving [TST07]’s work about robust estimation for texture flow. According to [BSW03] and [CM02], texture flow should have close relationship with texture intrinsic scale. Once we can compute the texture flow vector field, many things can be done: 1) compute the intrinsic relative scale for each pixel; 2) construct multi-level texture stack. Therefore, we can extend texture synthesis for perspective example to more general inputs, such as field guided textures or textures on more general surfaces.