Internships

281 – Programmer, Lightweight and Deployable 3D Human Performance Capture for Automultiscopic Virtual Humans

Project Name
Lightweight and Deployable 3D Human Performance Capture for Automultiscopic Virtual Humans

Project Description
The lab is developing a lightweight 3D human performance capture method that uses very few sensors to obtain a highly detailed, complete, watertight, and textured model of a subject (clothed human with props) which can be rendered properly from any angle in an immersive setting. Our recordings are performed in unconstrained environments and the system should be easily deployable. While we assume well-calibrated high-resolution cameras (e.g., GoPros), synchronized video streams (e.g., Raspberry Pi-based controls), and a well-lit environment, any existing passive multi-view stereo approach based on sparse cameras would significantly under perform dense ones due to challenging scene textures, lighting conditions, and backgrounds. Moreover, much less coverage of the body is possible when using small numbers of cameras.

Job Description
We propose a machine learning approach and address this challenge by posing 3D surface capture of human performances as an inference problem rather than a classic multi-view stereo task. Intern will work with researchers to demonstrate that massive amounts of 3D training data can infer visually compelling and realistic geometries and textures in unseen region. Our goal is to capture clothed subjects (uniformed soldiers, civilians, props and equipment, etc.), which results in an immense amount of appearance variation, as well as highly intricate garment folds.

Preferred Skills

  • C++, OpenGL, GPU programming
  • Experience with computer vision techniques: multi-camera stereo, optical flow, facial feature, detection, bilinear morphable models, texture synthesis, markov random field
  • Operating System: Windows

Apply now.

Go back to the summer research program list.