Multi-view Recognition Using Weighted View Selection

In this paper, we present an algorithm for multi-view recognition in a distributed camera setting that learns which viewpoints are most discriminative for particular instances of ambiguity. Our method is built on top of 2D recognition algorithms and casts view selection as the problem of optimizing kernel weights in multiple kernel learning. The main contribution is a locality-sensitive meta-training step to learn a disambiguation function to select the relative weighting of available viewpoints needed to classify a 2D input example. Our method outperforms related approaches on benchmark multi-view action recognition data sets.