Efficient Optimization for Rank-based Loss Functions

CVPR 2018 (oral), HM best paper award

Authors: Pritish Mohapatra*, Michal Rolinek*, C.V. Jawahar, Vladimir Kolmogorov, M. Pawan Kumar

Direct optimization of rank-based metrics can be done via structured SVM framework. But then there is a nasty algorithmical subproblem to deal with. We dealt with it in optimal runtime.

Remark from the future: The whole structured SVM approach has visible limitations for this problem. Compared to #blackboxbackprop it looks very unnatural. Take a look at newer paper to see how the same problem can be tackled more elegantly

Links: Arxiv