An efficient multi-resolution SVM network approach for object detection in aerial images

In this paper, we deal with the problem of object detection in aerial images. A lot of efficient approaches uses a cascade of classifiers which process vectors of descriptive features such as HOG. In order to take into account the variability in object dimension, features at different resolutions are often concatenated in a large descriptor vector. This prevents from taking into account explicitly the different resolutions but results in losing some valuable information. To overcome this problem, we propose to use a new method based on a SVM network. Each resolution is processed, regardless to the others, at the input layer level, by a dedicated SVM. The main drawback of using such a network is that the computational complexity for the classification phase drastically increases. We propose then to foster an incomplete exploration of the network by defining an activation path. This activation path determines an order to activate the network neurons, one after the other, and introduces a rejection rule which allows the process to end before crossing the whole network. Experimental results are obtained and assessed in an industrial application of urban object detection. We can observe an average gain of 17% in precision while the computational cost is divided by more than 5, with respect to a standard method.