1: \begin{abstract}

2:     Structural support vector machines (SSVMs) are amongst the 

3:     best performing models for structured computer vision tasks, 

4:     such as semantic image segmentation or human pose estimation. 

5: 	%

6:     Training SSVMs, however, is computationally costly, because it 

7:     requires repeated calls to a structured prediction subroutine 

8:     (called \emph{max-oracle}), which has to solve an optimization 

9:     problem itself, \eg a graph cut. 

10: 	

11:     In this work, we introduce a new algorithm for SSVM training that

12:     is more efficient than earlier techniques when the max-oracle is 

13:     computationally expensive, as it is frequently the case in computer

14:     vision tasks. The main idea is to (i) combine the recent stochastic 

15:     Block-Coordinate Frank-Wolfe algorithm with efficient hyperplane 

16:     caching, and (ii) use an automatic selection rule for deciding whether 

17:     to call the exact max-oracle or to rely on an approximate one based

18:     on the cached hyperplanes. 

19:     

20:     We show experimentally that this strategy leads to faster convergence 

21:     to the optimum with respect to the number of requires oracle calls, 

22:     and that this translates into faster convergence with respect to the 

23:     total runtime when the max-oracle is slow compared to the other steps 

24:     of the algorithm.

25: 

26:     A publicly available C++ implementation is provided.

27: \end{abstract}

28: