Stanford CS 231n: Convolutional Neural Networks for Visual Recognition

Class Projects

Marker-less Motion Capture with 2 Fisheye Cameras

For our class project we are re-implementing a technique based on a paper developed by Max Plank Institute for Informatics and Intel Visual Computing Insititute. The concept is to track, in real time, the 3D kinematics of the human body during arbitrary activities based on a portable sensing system using only 2 fish-eye cameras. Essentially this project aims to build off of an existing implementation to create accurate human motion capture systems without the expensive overhead infrastructure costs and limiting capture volume of traditional optical marker-based systems (Vicon, OptiTrack, etc.). Our final project paper can be read here.


Image Source: Helge Rhodin, Christian Richardt, Dan Casas, Eldar Insafutdinov, Mohammad Shafiei, Hans-Peter Seidel, Bernt Schiele, Christian Theobalt. "EgoCap: Egocentric Marker-less Motion Capture with Two Fisheye Cameras". SIGGRAPH Asia 2016