Project 1: Hybrid Images

Tala Huhe (thuhe)

Overview

We aim to create hybrid images using a modification of the algorithm described in this 2006 SIGGRAPH paper. These are images generated by overlaying the high-frequency components, or details, of one image onto the low-frequency components, or background, of another. Due to the way humans perceive detail, these static images will appear to change content when viewed from different distances.

Algorithm

The algorithm is fairly straightforward. Given two input images, we simply:

1. Construct the Gaussian-Laplacian pyramids for both images.
2. Pick a cutoff frequency. This is equivalent to choosing the largest N Laplacian pyramid levels from the high-pass image.
3. Build the hybrid Laplacian pyramid by taking the largest N levels from the high-pass image and the rest of levels from the low-pass.
4. Generate the hybrid image using the smallest Gaussian level from the low-pass image and the combined Laplacian pyramid.

Results

Results were highly dependent on scene composition. Overall, images with similar features (e.g. cat and dog) produced the best results. Outputs often suffered from either low-frequency noise (i.e. being too blurry) at regular size, or high-frequency noise (i.e. salt-and-pepper dots) when downsampled. Also, as expected, images with relatively few details (e.g. line drawings) were good low-pass images, and vice-versa.