Finding People in Internet Images

In this paper we propose a new contour based approach for people detection in Internet images. This approach relies on a large part on perceptual organization principles, at either low level edge detection and higher geometrical level. Our method is mainly compounded of two stages designed to work together: a first stage of contour detection, then a second stage of geometrical analysis. We first describe our new low level perceptually edge detector and we show its ability of finding robust edges in many types of images. We then examine the higher level of geometrical processing that we have designed to find bodies among edges extracted by the first stage. We show results obtained on an image database and images taken over Internet.