Motion magnification is a technique that reveals subtle, invisible movements in video – such as a building’s vibration or a person’s pulse – by computationally amplifying small pixel displacements. However, current methods require the camera to be perfectly stationary on a tripod; any camera shake gets amplified alongside the scene motion, producing severe artifacts. This project aims to develop a hand-held motion magnification algorithm that can separate camera-induced motion from real scene motion in videos captured from a moving camera. Our approach leverages the insight that handheld camera motion is predominantly rotational, meaning the induced pixel displacements can be modeled as homographies without requiring depth estimation. By estimating this camera-motion subspace and projecting it out, we can isolate and amplify only the scene-intrinsic motions. We will validate our method by comparing handheld results against tripod-based ground truth across diverse scenes, with the long-term goal of enabling real-time motion magnification on a smartphone.

I want to participate in SuperUROP because I want to explore the computer vision side of combining design and AI. After taking computer vision last semester, I want to go deeper into the computational imaging foundations that underlie so much of the visual technology I care about.