Section: New Results

Optimizing Average Precision

Paticipants: Puneet Kumar Dokania, Aseem Behl, Pritish Mohapatra, C.V. Jawahar, M. Pawan Kumar

Average precision (AP) is one of the most commonly used measures for ranking. However, due to the inefficiency of optimizing it during learning, a common approach is to use surrogate loss functions such as 0-1 loss. In [27] , we proposed a new optimization algorithm for AP-SVM that allows training in a similar time to binary SVM. In [23] , we extended the AP-SVM framework to score the samples according to high-order information, as opposed to simple first-order information used in prior work. Finally, in [19] , we proposed a novel latent AP-SVM formulation that allows learning from weakly supervised datasets. The advantage of learning with high order and missing information is demonstrated on challenging computer vision problems such as action classification and object detection using standard benchmark datasets.