Abstract
In this thesis, we present a class-independent objectness estimation method searching possible object locations in one image. Towards this goal, we combine superpixels with graph-based segmentation to efficiently generate candidate regions for object detection. Beyond segmentation, our approach allows each of the hypotheses to iteratively merge with its neighboring superpixels based on their similarity. Then we use normed edge feature %which successfully reduces the noises of complex background's texture,to describe the close-boundary characteristics of the bounding box for each hypothesis and measure the associated objectness afterward by a pre-trained classifier. To utilize the advantage of exhaustive search and avoid generating too many high-overlap hypotheses, we diversify each step of our approach and apply non-maximal suppression(NMS) to refine the hypotheses. Thereafter, by using the proposed algorithm, we obtain a small set of high-quality hypothesized object locations(77.4% Mean Average Best Overlap(MABO) and 94.4% detection rate(DR) for all the objects in PASCAL VOC 2007 test set). By using the proposed strategies, the performance is increased to 89.2% MABO and 99.4% DR.