N Heads Are Better Than One:Exploring Theoretical Performance Bounds of 3D Face Reconstruction Methods

We introduce ``N Heads Are Better Than One'', a novel approach for evaluating combinations of existing 3D face reconstruction methods. By calculating lower theoretical error bounds for method combinations on the NoW benchmark, we establish a robust set of new baselines for the task of 3D face reconstruction. Our work also provides a framework for assessing the potential of these aggregate `pseudo-foundation models,' which leverage strengths from multiple existing approaches. In doing so, we improve understanding of the performance of current methods and set targets for future foundation models to beat.