Homework 2: Eulerian Video Magnification

In this project, I implement the system of Eulerian Video Magnification (Wu et al., 2012) for revealing subtle changes in the world.

The workflow of my code procedes as follows

• Read video
• Convert RGB color to NTSC format
• Build Laplacian pyramid. (I used some library function from the matlabPyrTools toolbox.)
• Run the bandpass filter on the given Laplacian pyramid. (During the experiment, I had some trouble to use the "fft" function. So in the code, I used the "filtfilt" to do the filtering. I found it is almost efficient to the "fft".)
• Amplify the siginal.
• Reconstruct the video from the Laplacian pyramid. (I used some library function from the matlabPyrTools toolbox.)
• Add back the amplified siginal to the original video.
• Write the 4-D matrix into a video.

The main challenge I encounter is the tuning of various parameters. Many parameters affects the result of the algorithm. For instance, in the third experiment, I used two different parameters for chrom attenuation. And the second result looks very bad compared to the first result.

In the last two experiments, I applied this tool for amplifying the movement on flowers and human action resepctively. I found the result on the flower is not very good. The reason might be the non-cyclic movement of the flowers and it is thus hard to determine the parameter for the frequency.

For the video of human action, the algorithm did a good job in amplifying the movement of the torso. However, it creates a lot of artifacts for the hand. This is because this tool is more useful for the subtle action as opposed to the large movment of the hand.