1: \begin{abstract}

2: In this paper, we reconsider the clustering problem for image over-segmentation from a new perspective.

3: %

4: We propose a novel search algorithm named ``active search''  which 

5: explicitly considers neighboring continuity. Based on this search 

6: method, we design a back-and-forth traversal strategy and a 

7: ``joint'' assignment and update step to speed up the algorithm. 

8: Compared to earlier works, such as Simple Linear Iterative Clustering 

9: (SLIC) and its follow-ups, who use fixed search regions and perform the assignment

10: and the update step separately, our novel scheme reduces the 

11: %iteration number before convergence,

12: number of iterations required for convergence,

13: and also improves the boundary sensitivity of the over-segmentation results.

14: Extensive evaluations on the Berkeley segmentation benchmark 

15: verify that our method outperforms competing methods under various

16: evaluation metrics.

17: %

18: In particular, lowest time cost is reported among existing 

19: methods (approximately $30\mbox{ fps}$ for a $481 \times 321$ image

20: on a single CPU core).

21: %

22: To facilitate the development of over-segmentation, the code will be

23: \emph{publicly available}.

24: \end{abstract}

25: