Our
Data
For our project, we will be using a mixture of existing datasets used in previous studies and applications of this type of motion amplification, as well as additional datasets we collect and analyze ourselves.
​
The purpose of the existing datasets is twofold.
One: the data gives us an understanding of what the final output should look like, if we implement our system correctly.
Two: the data gives us a way to test our algorithm to see if it is working effectively compared to existing processed data.
​
In short, we need a way to validate our algorithm both by the output result but also in comparison to what has been done before. Without this, we risk not understanding our results when we input our own data, and risk the algorithm not working entirely because of a lack of testing.
The Input Data
One of the initial datasets we are working with is provided for us by the MIT research paper we are basing our project off of. This video features a baby laying in a crib, breathing. There is slight movement in the chest along the zipper, and a small movement by the hand that is almost imperceptible without assistance. Take a look below
Color Amplification
We began our video amplification journey by working on a color amplification scheme. The purpose of this was to begin identifying movement and amplifying the color around it. Not only does this allow us to begin building experience with the filtering and determining movement, but also the basics of amplification using Gaussian and Laplacian pyramids. These were important skills later used in our motion amplification method. Take a look at this color amplified video below.
Motion Amplification
Building on our experience with the color amplification, we implemented an Eulerian motion amplification algorithm. This involves the passing of pixel movement (compared frame by frame) into a bandpass filter, out of which come the values we want to amplify. These are then multiplied by the amplification factor, yielding the result below: