Patch-based natural object detection using CF*IRF

We propose a patch-based approach for detecting natural objects on keyframes of video shots. We apply it on the extraction of semantic feature "vegetation" and "animal", and on some search tasks in TRECVID2003. Our detection method is based on color and texture features, and considers the spatial information as well. TRECVID evaluation shows that our approach works effectively and can deal with the special situation when the target object only occupies a small portion of the whole image. The main contribution of our approach is as follows: first, we devise a novel color weighting scheme which is named CF*IRF. Second, we use a patch-based detection method for the feature extraction task, and test it in an open large video corpus. Finally, spatial constraints of patches are defined in image tessellation, which provides more flexibility.