Unified Method Achieves State-of-the-Art 4D Face Reconstruction from Image Sequences

Researchers have developed a novel method for high-fidelity 4D facial reconstruction from image sequences, addressing challenges posed by non-rigid deformations and viewpoint variations. This unified approach utilizes canonical facial point prediction to transform dense tracking and dynamic reconstruction into a canonical reconstruction problem, yielding temporally consistent geometry and reliable correspondences. The transformer-based model simultaneously predicts depth and canonical coordinates, trained on multi-view geometry data. Experimental results demonstrate that this method achieves approximately three times lower correspondence error and improves depth accuracy by 16%, setting a new standard in reconstruction and tracking tasks.