Stephen Gould (Australian National University)
Tuesday, 29.3.2016, 14:30
In this talk I will first give a brief overview of the topic of bi-level mathematical programming, in which the solution of an inner optimization problem is used within the objective function of a outer problem. These problems were originally studied in the context of two-player games but have been recently applied to computer vision and machine learning applications. I will then discuss current work in our group that applies bi-level optimization to the problem of activity recognition in videos. Here we train an end-to-end classifier consisting of a convolutional neural network to encode video frames and rank pooling operation to capture activity dynamics. Our method achieves state of the art results on a number of activity recognition benchmark datasets.