1: \begin{abstract}

2:     We present a convolutional neural network (CNN) based solution for

3:     modeling physically plausible spatially varying surface

4:     reflectance functions (SVBRDF) from a single photograph of a

5:     planar material sample under unknown natural

6:     illumination. Gathering a sufficiently large set of labeled

7:     training pairs consisting of photographs of SVBRDF samples and

8:     corresponding reflectance parameters, is a difficult and arduous

9:     process.  To reduce the amount of required labeled training data,

10:     we propose to leverage the appearance information embedded in

11:     unlabeled images of spatially varying materials to self-augment

12:     the training process. Starting from an initial approximative

13:     network obtained from a small set of labeled training pairs, we

14:     estimate provisional model parameters for each unlabeled training

15:     exemplar. Given this provisional reflectance estimate, we then

16:     synthesize a novel temporary \emph{labeled} training pair by

17:     rendering the exact corresponding image under a new lighting

18:     condition. After refining the network using these additional

19:     training samples, we re-estimate the provisional model parameters

20:     for the unlabeled data and repeat the self-augmentation process

21:     until convergence.  We demonstrate the efficacy of the proposed

22:     network structure on spatially varying wood, metals, and plastics,

23:     as well as thoroughly validate the effectiveness of the

24:     self-augmentation training process.

25:     

26: \end{abstract}