1: \begin{abstract}

2:       Estimating shape and appearance of a three dimensional object from a given

3:       set of images is a classic research topic that is still actively pursued.

4:       Among the various techniques available, \glsentrylong{PS} is distinguished

5:       by the assumption that the underlying input images are taken from the same

6:       point of view but under different lighting conditions. The most common

7:       techniques provide the shape information in terms of surface normals. In

8:       this work, we instead propose to minimise a much more natural objective 

9:       function, namely the reprojection error in terms of depth.

10:       Minimising the resulting non-trivial variational model for

11:       \glsentrylong{PS} allows to recover the depth of the photographed

12:       scene directly. As a solving strategy, we follow an approach based on a

13:       recently published optimisation scheme for non-convex and non-smooth cost functions.\par 

14:       %

15:       The main contributions of our paper are of theoretical nature. A

16:       technical novelty in our framework is the usage of matrix differential

17:       calculus. We supplement our approach by a detailed convergence analysis of the resulting

18:       optimisation algorithm and discuss possibilities to ease the computational 

19:       complexity. At hand of an experimental evaluation we

20:       discuss important properties of the method. Overall, our strategy 

21:       achieves more accurate results than competing

22:       approaches. The experiments also highlights some practical aspects of the 

23:       underlying optimisation algorithm 

24:       that may be of interest in a more general context. 

25: \end{abstract}