1: \begin{abstract}

2: Learning a fast and discriminative patch descriptor is a challenging topic in 

3: computer vision. Recently, many existing works focus on training various descriptor 

4: learning networks by minimizing a triplet loss (or its variants), which is expected 

5: to decrease the distance between each positive pair and increase the distance between 

6: each negative pair. However, such an expectation has to be lowered due to the 

7: non-perfect convergence of network optimizer to a local solution. Addressing this 

8: problem and the open computational speed problem, we propose a 

9: \textit{Des}criptor \textit{Dis}tillation framework for local descriptor learning, 

10: called DesDis, where a student model gains knowledge from a pre-trained teacher

11: model, and it is further enhanced via a designed teacher-student regularizer. 

12: This teacher-student regularizer is to constrain the difference between the positive 

13: (also negative) pair similarity from the teacher model and that from the student 

14: model, and we theoretically prove that a more effective student model could be 

15: trained by minimizing a weighted combination of the triplet loss and this regularizer, 

16: than its teacher which is trained by minimizing the 

17: triplet loss singly. Under the proposed DesDis, many existing descriptor networks 

18: could be embedded as the teacher model, and accordingly, both equal-weight and 

19: light-weight student models could be derived, which outperform their teacher in 

20: either accuracy or speed. Experimental results on 3 public datasets demonstrate 

21: that the equal-weight student models, derived from the proposed DesDis framework 

22: by utilizing three typical descriptor learning networks as teacher models, could 

23: achieve significantly better performances than their teachers and several other 

24: comparative methods. In addition, the derived light-weight models could achieve 

25: 8 times or even faster speeds than the comparative methods under similar patch 

26: verification performances.

27: 

28: \end{abstract}

29: