SI-Cut: Structural Inconsistency Analysis
for Image Foreground Extraction



I-Chen Lin, Yu-Chien Lan, Po-Wen Cheng
Corresponding author: I-Chen Lin


This paper presents a novel approach for extracting foreground objects from an image. Existing methods involve separating the foreground and background mainly according to their color distributions and neighbor similarities. This paper proposes using a more discriminative strategy, structural inconsistency analysis, in which the localities of color and texture are considered. Given an indicated rectangle, the proposed system iteratively maximizes the consensus regions between the original image and predicted structures from the known background. The object contour can then be extracted according to inconsistency in the predicted background and foreground structures. The proposed method includes an efficient image completion technique for structural prediction. The results of experiments showed that the extraction accuracy of the proposed method is higher than that of related methods for structural scenes, and is also comparable to that of related methods for less structural situations.


The proposed method was compared with three related state-of-the-art methods: GrabCut [Rother et al. 2004], the pinpoint method with a bounding box prior [Lempitsky et al. 2009] (abbreviated as Box-prior), and the GrabCut in one cut [Tang et al. 2013] (abbreviated as One-cut).

Several segmentation results of the proposed and comparative methods are shown below. Please refer to the manuscript and  supplementary files for more results and comparison.

Figure. Comparative results for the structural scene dataset. From left to right: Input images and rectangles, GrabCut results, Box-prior results, One-cut results (with individually optimal weights), the proposed results (AE: with auto-estimation iterations), the proposed results (UA: with user-assigned iterations).

Figure. Comparative results for the GrabCut dataset. From left to right: Input images and rectangles, Grab- Cut results (from [Lempitsky et al. 2009]), Box-prior results [Lempitsky et al. 2009], One-cut results [Tang et al. 2013], the proposed (AE) results, the proposed (UA) results.

Datasets used in the experiments

Structural scene dataset (SSDB40)
The 40 source images were selected from the LabelMe project page. The credits for the image sources go to the LabeMe database [Russel et al. 2008]. A large portion of ground-truth masks in LabelMe are approximated by polygonal contours. For more accurate experiments, the ground-truth masks were refined by users and used in this experiment.
Download: (zipped file, about 16.7MB, including images, indicated rectangles and ground-truth masks)
                     SSDB40_src_mapping (The mappings between images in SSDB40 and LabelMe are listed in this table)

GrabCut dataset
The 50 source images and ground truth masks of GrabCut dataset are from GrabCut project page [Rother et al. 2004].
The indicated rectangles used in this experiment are identical to those used in [Lempitsky et al. 2009].

(NOTE: As emphasized in [Lempitsky et al. 2009], the results and error rates reported here are based on bounding-rectangle inputs. They are not appropriate for comparison with results and error rates of related methods based on the trimap, lasso or scribble inputs.)


I-Chen Lin, Yu-Chien Lan, Po-Wen Cheng, "SI-Cut: Structural Inconsistency Analysis for Image Foreground Extraction," IEEE Trans. Visualization and Computer Graphics, 21(7):860-872, July, 2015.

Paper: preprint_version (about 19.8MB), published version (link to the IEEE digital library)
Supplementary file: TVCG15_sup_results (pdf, about 16.7MB)


author = {I-Chen Lin and Yu-Chien Lan and Po-Wen Cheng},
title = {{SI-Cut}: Structural Inconsistency Analysis for Image Foreground Extraction},
journal = {{IEEE} Transactions on Visualization and Computer Graphics},
volume = {21},
number = {7},
pages = {860--872},
month = {July},
year = {2015},
doi = {10.1109/TVCG.2015.2396063}


The authors appreciate the helpful comments from the anonymous reviewers. This paper was partially supported by the Ministry of Science and Technology, Taiwan under grant no. MOST 103-2221-E-009-143.

Go back to I-Chen Lin's publication webpage (English)