PURE sharpens 3D reconstruction from unlabeled video

PURE targets the biggest weakness in current feed-forward reconstruction: brittleness around motion, occlusion, and changing illumination. Its practical value lies in adapting proven backbones rather than introducing another heavyweight reconstruction pipeline.

• –Uses relative depth, local geometry, and filtered feature correspondences as pseudo-supervision
• –Trains on millions of real unlabeled frames while mixing synthetic data to reduce catastrophic forgetting
• –Preserves the original models’ inference cost, making deployment more practical
• –Open-source code, weights, and a hosted demo lower the barrier for experimentation
• –Results still depend on teacher-model quality and may retain seams or motion artifacts in difficult scenes