Efficient objectness via saliency seeds and contour segments

Object proposal is a new paradigm for improving efficiency for object detection. We propose an efficient method for object proposals by saliency seeds and contour segments. A simple saliency method is used to get several salient seeds in the image to target all the probable objects appeared in image, roughly leaving background regions out of consideration. Then we further score each of the salient seeds by using a bounding box strategy. If the bounding box contains more contour segments of the seed, it is assumed to be the object proposal more strongly. For efficiency, we utilize Pair of Adjacent Segments (PAS) as the contour segment feature, which is easy to detect and can describe the location and scale of contours compactly. After getting the proposal regions, those PAS features are also used for classification task. Experiments show that the proposed method is very effective. It has achieved comparable result to state of the art methods with higher efficiency and also provide auxiliary information to later classification step.