Capturing the shape and appearance of a human has been a significant first step for applications such as clothing fitting, avatar creation, and fitness, to name a few. With the release of powerful, relatively low-cost cameras on mobile devices, there are new opportunities for achieving this task. In this project, the goal is to investigate recent deep neural network architectures for the reconstruction of human shape and texture. Given a set of images (RGB or RGBD) of a human taken from multiple views, the objective is to reconstruct the geometry of the human body and a highly detailed texture for the mesh of the human body. Methods for segmenting the body into parts or detecting body landmarks will also be investigated.
Basic knowledge of computer vision and deep learning. Programming skills: python, TensorFlow (optional), PyTorch (optional).
The project involves collecting the relevant literature on human body reconstruction, testing existing neural network architectures, comparing and reporting on their results, and considering how their performance can be improved.
Expected deliverables: Final report, code base