Light Field Surface Feature and Spherical Descriptor in the Surface Scale Space

Depth information calculated from the spatial-angular consistent light field data of a Lambertian scene without occlusion provides a Monge surface representation of the scene surface. The light field surface gradient and Laplace operators can then be defined on the moving frame. The light field surface scale space can be established by solving the diffusion equation, and the light field surface feature (LFSF) can be detected. Then the local surface gradients of each LFSF are measured and form a 48 dimensional rotation invariant spherical descriptor. Experimental results show that the proposed method can detect more features then LiFF and Lisad on most of the scenes. For light field data matching, the proposed method provides more matching pairs than LiFF with SIFT descriptor and procides more accurate matching result.