Gradient domain methods with application to 4D scene reconstruction

• We present a generalization of some integration schemes for dynamic scenes.
• The presented framework can improve Differential 3D shape retrieval methods.
• We give implementation details and put the code publicly available.
• Different regularization terms are proposed and discussed.

In many applications such as Photometric Stereo, Shape from Shading, Differential 3D reconstruction and Image Editing in gradient domain it is important to integrate a retrieved gradient field. In most of the real experiments, the retrieved gradient fields correspond to nonintegrable fields (i.e. they are not irrotational on every point of the domain). Robust approaches have been proposed to deal with noisy nonintegrable gradient fields. In this work we extend some of these techniques for the case of dynamic scenes when the gradient field in the x−yx−y domain can be estimated over time. We exploit temporal consistency in the scene to ensure integrability and improve the accuracy of the results. In addition, two known integration algorithms are reviewed and important implementation details are discussed. Experiments with synthetic and real data showing some potential applications for the proposed framework are presented.