Sub-scene generation: A step towards complex scene understanding

This paper proposes a new method for generating sub-scenes from images of natural scenes. It is based on over-segmented image patches initially and uses rules derived from human psychology to merge the patches into reasonable sub-scenes. This merging process focuses on covering semantic gaps by exploring semantic connections among the patches, such as the influence area of each patch where the potential interaction with other patches may exist; and applying visual organization rules of proximity property and color harmony property. The proposed method is simple yet efficient computationally. It does not have any limitation on the input image class and does not include a training process. Our experiments indicate that it handles complex scenes with clutters and occlusions very well, and outperforms general segmentation methods by its meaningful semantic grouping and self determined stop criteria.