발표자: 이세호(고려대 박사과정)
발표일: 2017.8.
개요:
슈퍼픽셀 알고리즘은 입력 영상을 다수의 의미 있는 영역으로 과분할 하는 기법이다. 입력 영상을 픽셀 단위로 표현할 때와 비교하여, 슈퍼픽셀 단위의 표현은 입력 영상의 단위의 수를 크게 줄이는 장점이 있다. 각 슈퍼픽셀은 객체의 윤곽선을 넘어서는 영역을 포함하지 않는 동시에, 단일 객체만을 담아야 한다. 본 발표에서는 객체의 윤곽선 정보를 고려한 윤곽선 제약 슈퍼픽셀 기법(contour-constrained superpixel algorithm)을 제안한다.
3. • Superpixel
• Over-segment input images into a set of meaningful regions
• Can reduce the number of image primitives or units greatly
• Superpixels should adhere to object contours
Introduction
5. Superpixel
• Conventional methods
• K-means based methods
• Assign each pixel to the nearest cluster and update the cluster
center iteratively
• SLIC (TPAMI’12), LSC (CVPR’15), and MSLIC (CVPR’16)
• Coarse-to-fine methods
• Change the superpixel label of boundary regions iteratively based
on the cost function
• From block-level to pixel-level
• SEEDS (IJCV’15) and Yao et al. (CVPR’15)
6. Contour-Constrained Superpixels
• Contributions
• Introduction of the contour constraint
• Extension of the proposed algorithm for video processing
• Remarkable performance achievement and improvement of
many computer vision algorithms
8. Contour-Constrained Superpixels
• Superpixel labeling
• Hierarchical block structure
• From block to pixel levels
• Divide only inhomogeneous regions
• Can maintain relatively regular and compact shape
9. Contour-Constrained Superpixels
• Superpixel labeling
• Change the label of region (block or pixel) 𝑅𝑖 from 𝑙 𝑅𝑖 to
𝑙 𝑅𝑗 of neighboring region 𝑅𝑗 ∈ 𝒩𝑅 𝑖
• Change the label of region by minimizing the cost function
• 𝐸 𝑖, 𝑗 = 𝐸D 𝑖, 𝑗 + 𝛾𝐸L 𝑖, 𝑗 + 𝜂𝐸I 𝑖, 𝑗 × 𝐸C 𝑖, 𝑗
• To preserve the topology of superpixels
• Check the boundary region 𝑅𝑖 is simple point
Initial Level 1 Level 2 Level 3 Level 4
10. Superpixel Labeling
• Feature distance from superpixel centroid
• 𝐸D 𝑖, 𝑗 = 𝐜 𝑅𝑖 − 𝐜 𝑆𝑙 𝑅 𝑗
2
+ 𝐩 𝑅𝑖 − 𝐩 𝑆𝑙 𝑅 𝑗
2
• Color distance term makes the color of each superpixel to be regular
• Spatial distance term imposes the superpixels to be distributed
compactly
Color distance Spatial distance
11. Superpixel Labeling
• Boundary length cost
• 𝐸L 𝑖, 𝑗 = 𝜆 𝑅𝑖, 𝑙 𝑅𝑗 − 𝜆 𝑅𝑖, 𝑙 𝑅𝑖
• 𝜆 𝑅𝑖, 𝑘 : total number of boundary regions when the superpixel
label of 𝑅𝑖 is 𝑘
• Minimize the boundary length of superpixels
• Make the superpixels distributed compactly
12. Superpixel Labeling
• Inter-region color cost
• 𝐸I 𝑖, 𝑗 = max 0, 𝐜 𝑅𝑖 − 𝐜 𝑅𝑗
2
− 𝜅 𝑆𝑙 𝑅 𝑗
• Neighboring regions with dissimilar color information
→ Assign different superpixel labels
• Internal difference
• Measure the amount of texture information of superpixels
• 𝜅 𝑆𝑙 𝑅 𝑖
= max
𝑅 𝑚,𝑅 𝑛∈𝑆𝑙 𝑅 𝑖
,
𝑅 𝑛∈𝒩 𝑅 𝑚
𝐜 𝑅 𝑚 − 𝐜 𝑅 𝑛
2
• Maximum color difference between neighboring regions within 𝑆𝑙 𝑅 𝑖
13. Superpixel Labeling
• Contour constraint 𝐸C 𝑖, 𝑗
• Amplify the cost function when there is an object contour
between two regions
• Use holistically-nested edge detection (HED)*
• Deep learning based edge detection method
𝐸 𝑖, 𝑗 = 𝐸D 𝑖, 𝑗 + 𝛾𝐸L 𝑖, 𝑗 + 𝜂𝐸I 𝑖, 𝑗 × 𝐸C 𝑖, 𝑗
* S. Xie and Z. Tu, Holistically-nested edge detection. In ICCV, pages 1395-1403, 2015.
14. Contour Constraint
• Contour pattern set extraction
• Use 200 training images in the BSDS500 dataset
• Denote patch centered at each contour pixel as a contour
pattern
• Select top 1,000 frequently occurring patterns
15. Contour Constraint
• Contour pattern set
• Only consider the patterns which divide the patch into two regions
• 1,000 patterns cover 90.5% of the patches in the training dataset
16. Contour Constraint
• Contour pattern matching
• Get thin binary contour map
• Perform non-maximum suppression and then do thresholding
• Compute Hamming distances from the contour patterns and
select the best matching pattern
17. Contour Constraint
• Contour probability 𝜙 𝑢, 𝑣
• Difficult to estimate the existence of object contours only
considering the thin contour map
• Measure the proportion of patches whose matching patterns
separate 𝑢 from 𝑣
• Consider patches containing both pixels 𝑢 and 𝑣
18. Contour Constraint
• Determining contour constraint 𝐸C 𝑖, 𝑗
• Contour probability 𝜓 𝑅𝑖, 𝑅𝑗 between regions
• Find the maximum contour probability between the pixels in 𝑅𝑖
and 𝑅𝑗
• 𝜓 𝑅𝑖, 𝑅𝑗 = max
𝑢∈𝑅 𝑖,𝑣∈𝑅 𝑗
𝜙 𝑢, 𝑣
• 𝐸C 𝑖, 𝑗 = exp 𝛽 × 𝜓 𝑅𝑖, 𝑅𝑗
19. Hierarchical Superpixel Refinement
• Hierarchical block structure
• Divide only inhomogeneous regions into four blocks
• Dissimilarity function
• 𝜃 𝑅𝑖 = max
𝑢,𝑣∈𝑅 𝑖,𝑣∈𝒩𝑢
𝐜 𝑢 − 𝐜 𝑣 2 + exp 𝛽 × max
𝑢,𝑣∈𝑅 𝑖,𝑣∈𝒩𝑢
𝜙 𝑢, 𝑣
• Divide the region 𝑅𝑖 if 𝜃 𝑅𝑖 > 𝜏div
22. Temporal Superpixels
• Initialization
• To make temporally consistent superpixels
• Estimate optical flows from 𝐼 𝑡−1
to 𝐼 𝑡
• Transfer the label of each superpixel by employing the average
optical flow of the superpixel
• Do not assign any labels to occluded or disoccluded pixels
• Occluded pixel: a pixel mapped from multiple superpixels
• Disoccluded pixel: a pixel mapped from no superpixel
23. Temporal Superpixels
• Temporal superpixel labeling
• Performed at the pixel level only
• Based on the energy function
• 𝐸 𝑖, 𝑗 = 𝐸D 𝑖, 𝑗 + 𝛾𝐸L 𝑖, 𝑗 + 𝜂𝐸I 𝑖, 𝑗 × 𝐸T 𝑖, 𝑗
• 𝐸T 𝑖, 𝑗 : temporal contour constraint
24. Temporal Superpixel Labeling
• Temporal Contour constraint
• Make superpixels temporally consistent
• Make superpixels compatible with image contours
• 𝐸T 𝑖, 𝑗, 𝑡 = 𝐸C 𝑖, 𝑗, 𝑡 × 𝜌 𝑖, 𝑗, 𝑡
• Relaxation factor 𝜌 𝑖, 𝑗, 𝑡
• 𝜌 𝑖, 𝑗, 𝑡 = ൞
1
1+exp −𝜁×ℎ 𝑅𝑖
𝑡
, if 𝑙 𝑅𝑗
𝑡
∈ ℒ 𝑖
𝑡
1, otherwise
• ℒ 𝑖
𝑡
is the set of labels that are mapped to 𝑅𝑖
𝑡
from 𝐼 𝑡−1
• Relax 𝐸T 𝑖, 𝑗, 𝑡 when 𝑅𝑖
𝑡
does not contain object contour
25. Merging, Splitting, and Relabeling
• Merging and splitting
• Prevent irregular superpixel size
• When Τ𝐴 𝑘
𝑡 ҧ𝐴 is larger than 𝜏spl → split
• Consider the biggest eigenvector of the spatial distribution
• When Τ𝐴 𝑘
𝑡 ҧ𝐴 is smaller than 𝜏mer → merge
• Merge to the nearest superpixel
26. Merging, Splitting, and Relabeling
• Relabeling
• Avoid incorrect labeling
• Because of occlusion or illumination variation
• Define color consistency 𝐶 𝑘
• 𝐶 𝑘 = 𝐜1:𝑡−1 𝑆 𝑘 − 𝐜 𝑡 𝑆 𝑘
2
• If 𝐶 𝑘 is larger than 𝜏rel → relabel
33. Experimental Results
• Application to video object segmentation
• To superpixel-based video object segmentation method*
• Use CCS instead of SLIC as a preprocessing
• Intersection over union (IoU) is increased from 0.532 to 0.571
* W.-D. Jang and C.-S. Kim, Semi-supervised video object segmentation using multiple random walkers. In BMVC, 2016.
34. Experimental Results
• Application to video saliency detection
• Postprocessing
• Average the saliency values of the pixels in all frames, constituting
each superpixel
• Applied to HS* and DHSNet**
* Q. Yan, L. Xu, J. Shi, and J. Jia, Hierarchical saliency detection. In CVPR, 2013.
** N. Liu and J. Han, DHSNet: Deep hierarchical saliency network for salient object detection. In CVPR, 2016.
36. Conclusions
• Contour-constrained superpixels (CCS)
• Perform hierarchical refinement from block to pixel levels
• Based on the contour constraint
• Temporal superpixels
• CCS algorithm for video processing
• Use optical flow to obtain temporally consistent superpixels
• Experimental results
• CCS outperforms the state-of-the-art superpixel methods
• Can be applied to object segmentation and saliency detection