Multi-view Geometry Estimation for Light Field Compression
Ismail Oner Sebe, Prashant Ramanathan, Bernd Girod · 2002
Geometry-aided light field compression requires accurate geometry for the efficient representation of the light field image data. In this work, we propose a method to directly estimate a geometry model from the light field images, such that it maximizes the compression efficiency. Previous work on this problem uses least-squares to solve a set of optical-flow based equations. This approach suffers from stability problems and gives sub-optimal results, due to outliers included in the set of equations. We propose an extension to this approach that addresses some of its short-comings. Specifically, we use weighted least-squares to identify and suppress the effects of outliers and a multi-resolution estimation algorithm for the geometry model. The experiments performed on real and synthetic data show that our technique stabilizes the algorithm, decreases the total running time by a factor of 10 and decreases bit-rate by up to 10% over the previous work. For the synthetic light field sequences, we show that this achieves the optimal achievable compression efficiency, under our constrained arrangement. 1