Sequential Attention Planning for Large-scale Environment Classification

In this paper, a sequential attention planning algorithm is proposed and applied to wide-uncertain environment classification with small field-of-view cameras. Attention planning is formulated as the sequential feature selection problem that greedily finds a sequence of attentions to obtain more informative observations, yielding faster training and higher accuracies. However, greedy algorithm cannot always guarantee the optimal solution. In order to find the near-optimal solution for attention planning, adaptive submodular optimization is considered under certain assumptions, where the objective function for the internal belief is adaptive submodular and adaptive monotone. First, the amount of information of individual attention area is modeled as dissimilarity variance among the environment data set, respectively. With this model, the information gain function is defined as a function of variance reduction that has been shown to be submodular and monotone in many cases. Furthermore, adapting to increasing numbers of observations, each information gain for attention areas is iteratively updated by discarding the non-informative prior knowledge about current environment using a cascade of nearest neighbor classifiers, enabling to maximize the expected information gain. The effectiveness of the proposed algorithm is verified through experiments that can significantly enhance the uncertain environment classification accuracy, with reduced number of limited field-of-view observations