Abstract
This paper presents a novel system for the 3D capture of facial performance using standard video and lighting equipment. The mesh of an actor's face is tracked non-sequentially throughout a performance using multi-view image sequences. The minimum spanning tree calculated in expression dissimilarity space defines the traversal of the sequences optimal with respect to error accumulation. A robust patch-based frame-to-frame surface alignment combined with the optimal traversal significantly reduces drift compared to previous sequential techniques. Multi-path temporal fusion resolves inconsistencies between different alignment paths and yields a final mesh sequence which is temporally consistent. The surface tracking framework is coupled with photometric stereo using colour lights which captures metrically correct skin geometry. High-detail UV normal maps corrected for shadow and bias artefacts augment the temporally consistent mesh sequence. Evaluation on challenging performances by several actors demonstrates the acquisition of subtle skin dynamics and minimal drift over long sequences. A quantitative comparison to a state-of-the-art system shows similar quality of temporal alignment. © 2012 IEEE.