Depth Estimation of Semi-submerged Objects Using a Light Field Camera

Abstract

We present a new algorithm to estimate depth of real-world scenes containing an object semi-submerged in water using a light field camera. Existing hand-held consumer light field cameras are well-suited for automated refocusing, depth detection in outdoor environment. However, when it comes to surveying marine environment and near water macro photography, all depth estimation algorithms based on traditional perspective camera model will fail because of the refracted rays. In this thesis, a new method is presented that explicitly accommodates the effect of refraction and resolves correct depths of underwater scene points. In particular, a semi-submerged object with opaque Lambertian surface with repeating textures is assumed. After removing the effect of refraction, the reconstructed underwater part of the semi-submerged object has consistent depth and shape with that of the above-water part. With experiments on synthetic scenes rendered with modeling software Blender and real light field images captured by a Lytro Illum camera, we show that our algorithm can largely remove the effect of refraction for semi-submerged objects using an image from a light field camera.