Research Project Title:
Analyzing 3D Reconstruction of Cameras and Spatial Information in Movies and Building a Deep Learning Model to Improve Current Algorithms
abstract:Extracting 3D information from 2D has been a growing focus. With recent progresses and more 3D datasets made available, it is possible to extract more complicated 3D information. Our research will focus on 3D reconstruction from movies using methods like structure from motion, stereo, and orbslam. Movie provides diverse and complicated information, difficult for machine to understand. We want to start with extracting 3D shape and location of objects from single frames, then use sequential frames to 3D reconstruct. And because movies focus mostly on people, we can use trajectory, face or pose to constrain the base 3D reconstruction. This project can contribute to creating VR simulations, reconstructing heritage and helping machine to learn from the immensity of video data out there.
I have been working on this project in the past year. The amount of real world applications our research can unlock always make me feel excited. Through participating in SuperUROP, I hope to strengthen skills like writing papers and doing presentation. I also hope to fill in knowledge gaps like computational geometry and computer vision algorithms.