Date: November 2011.
Source: 2011 IEEE International Conference on Computer Vision (ICCV), Barcelona, Spain, pp. 2296–2303. DOI: 10.1109/ICCV.2011.6126510
Research Summary: D3DFACS addresses a gap in 3D facial analysis research, the lack of dynamic 3D facial datasets coded to the Facial Action Coding System (FACS), the standard framework for describing facial muscle movement. The dataset captures 10 subjects, including 4 trained FACS experts, performing 519 real-world dynamic Action Unit (AU) sequences, with each sequence’s peak frame manually FACS coded by certified experts. The paper also introduces the first framework for building dynamic 3D morphable models, using a novel Active Appearance Model (AAM) based registration approach that avoids the drift problems common to optical flow methods, providing more reliable mesh correspondence across a facial motion sequence.
3dMD’s Role: D3DFACS’s dynamic facial sequences were captured on a 3dMD dynamic 3D stereo camera system, six cameras split across two pods, sampling at 60 frames per second and producing real-world 3D meshes of roughly 30,000 vertices with corresponding UV color texture. This is one of the earliest confirmed 3dMD-based datasets in this citation series, and it went on to train the expression space behind FLAME, directly cited in FLAME’s own methods.
Article: A FACS Valid 3D Dynamic Action Unit Database with Applications to 3D Dynamic Morphable Facial Modeling.
Authors: D Cosker, E Krumhuber, A Hilton. University of Surrey.