Efficient camera smoothing in sequential structure-from-motion using approximate cross-validation

Abstract

In the sequential approach to three-dimensional reconstruction, adding prior knowledge about camera pose improves reconstruction accuracy. We add a smoothing penalty on the camera trajectory. The smoothing parameter, usually fixed by trial and error, is automatically estimated using Cross-Validation. This technique is extremely expensive in its basic form. We derive Gauss-Newton Cross-Validation, which closely approximates Cross-Validation, while being much cheaper to compute. The method is substantiated by experimental results on synthetic and real data. They show that it improves accuracy and stability in the reconstruction process, preventing several failure cases. © 2008 Springer Berlin Heidelberg.