CS 194-26 Project #3: Fun with Frequencies and Gradients!

Overview

The purpose of this project was to explore different approaches to image blending.

Part 1: Frequency Domain

Part 1.1: Warmup

Here is a cat picture I captured on holiday.

Part 1.2: Hybrid Images

Watch out! You can't blend just any old Derek and Nutmeg!

The correct Derek and Nutmeg in the frequency domain:

Part 1.3: Gaussian and Laplacian Stacks

Lincoln and Gala, and Mona Lisa decomposed:

Additional examples:

Part 1.4: Multiresolution Blending

In this section, each laplacian layer is blended at a different intensity, with the help of a gaussian masking stack. One failure case is combining faces of different skin tones.

Part 2: Gradient Domain Fushion

In these examples, the goal is to transplant an object from one background to another. The object is masked by hand, and adjusted such that it blends into the new background. The adjusted object is computed to minimize the difference in gradient from the original object, and to minimize the difference from border pixels. This often results in a shift of hue.

Part 2.1: Toy Problem

Part 2.2: Poisson Blending

When we try to add parrot friends to the whale shark's house, they appear too dark because the hue of the addition image is shifted to merge the sky and water. The overall effect improves slightly when the sky is made darker, but the blend is not smooth.

Here are some images that worked a little bit better:

A comparison of techniques: