Robust Extraction of Optic Flow Differentials for Surface Reconstruction

The first-order differential invariants of optic flow, namely divergence, curl, and deformation, provide useful shape indicators of objects passing through view. However, as differential quantities these are often difficult to extract reliably. In this paper we present a filter-based method for computing these invariants with sufficient accuracy to permit the construction of a partial scene model. The noise robustness of our method is analysed using both synthetic and real world images. We also demonstrate that the deformation of a dense optic flow field encodes sufficient information to reliably estimate surface orientations if viewer ego-motion is purely translational.