Interesting region detection in aerial video using Bayesian topic models

Searching interesting regions in aerial video is a new and challenging problem. This paper presents an approach to detect visual interesting regions in aerial video using pLSA topic model. Traditional interesting region detection approaches just use bottom-up information, such as color, orientation and movement etc. Our proposed method can discover the semantic content of the whole image, the co-occurrence of local image patches via pLSA model, and consequently improve detection result significantly in real world scenes. First, we extract frames from aerial video as documents. Then we use vector quantized SIFT descriptors as words. Third, we discover topics (e.g. plants, roads, buildings) and the relation among them using pLSA model. Finally, we can detect interesting regions as we need according to calculated models. Experimental observations show the success of our approach on interesting region detection in aerial video.