A framework for designing natural object detector in video

This paper presents a framework for detecting natural objects in video. We analyze each key frame of video to determine if there is any interesting natural object inside. After key frames are extracted by our shot boundary detection (SBD) system, natural object detection in a key frame is done through following such a general pipeline. First, a novel representation CF*IRF (color frequency and inverse region frequency) is developed to detect all possible regions of the target object. Then those regions are validated by texture features and other color features. Finally the shape and location information of the regions give the final answer of the present of the target object. In this paper we also discuss the problem of search task's evaluation. A set of criteria is proposed, by which different search tasks can be classified and analyzed.