발표자: 이준태 (고려대 박사과정)
발표일: 2017.9.
개요:
Algorithm to semantic line detection and its applications will be presented. First, I will introduce the concept of semantic line. Second, semantic line detection method will be described. Then, two applications will be presented: composition enhancement and image simplification.
7. Semantic Line Network (SLNet)
• Line pooling layer
• Extract a regional representation of a candidate line from a
convolutional feature map
Line
Convolution activation Line pooling activation
9. Semantic Line Network (SLNet)
• Network training
• Learning rate 𝜖 ← 0.1𝜖 after every two epochs (start: 0.001)
• Momentum 0.9, weight decay 0.0005
• Input image size: 400 x 400
• Candidate lines
Ground truth Uniform sampling Augment positives
10. Primary Semantic Line
• Primary semantic line characterizes the photographic
composition of an image
• For 1,308 candidate lines, the candidate 𝐥max with the
maximum soft-max probability is selected
ሚ𝐥max = 𝐥max + ∆𝐥max
Initial candidate
Regressed line
11. Multiple Semantic Line
• A candidate line is declared as a semantic line and
regressed, if its soft-max probability is higher than 0.5
Input image Multiple semantic lines
12. Multiple Semantic Line
• Non-maximum suppression(NMS)
• Prevents semantic lines from being too close to one another
• Edge density for each regressed line ሚ𝐥
𝜌 ሚ𝐥 =
𝑁edge
𝑁all
A semantic line example Edge map [1]
[1] S. Xie, and T. Zhuowen, "Holistically-nested edge detection," in Proc. of the IEEE ICCV. 2015.
13. Multiple Semantic Line
• Non-maximum suppression(NMS)
• Prevents semantic lines from being too close to one another
• Measure mean intersection over union (mIoU) between two
semantic lines
• If mIoU > 0.85, two semantic lines are regarded as duplicated, and
the line with a lower edge density is removed
𝑅
𝑄
𝑅′
𝑄′
mIoU =
1
2
(
𝑅 ∩ 𝑅′
𝑅 ∪ 𝑅′
+
𝑄 ∩ 𝑄′
𝑄 ∪ 𝑄′
)
14. Semantic Line Detection Results
• Primary semantic line detection
• 𝑁𝑐 : the number of the test images whose primary semantic lines
are correctly detected
Accuracy =
𝑁𝑐
𝑁
• Multiple semantic line detection
• 𝑁𝑠 : the number of correctly detected semantic lines
• 𝑁𝑒 : the number of false positives
• 𝑁 𝑚 : the number of false negatives
Precision =
𝑁𝑠
𝑁𝑠 + 𝑁𝑒
, Recall =
𝑁𝑠
𝑁𝑠 + 𝑁 𝑚
15. Semantic Line Detection Results
• The accuracy, precision, and recall curves
• Area under curve scores
19. Application Ⅰ. Composition Enhancement
• Horizon estimation + aesthetic cropping
Primary semantic line Rotated image Cropping result
20. Application Ⅰ. Composition Enhancement
• Horizon estimation + aesthetic cropping
Primary semantic line Rotated image &
Aesthetic region
Cropping result
21. Application Ⅰ. Composition Enhancement
• Horizon estimation + aesthetic cropping
Primary semantic line Rotated image &
Aesthetic region
Cropping result
22. Application Ⅱ. Image Simplification
• Image simplification
• Divide an image into polygonal regions along semantic lines
• Make the regions homogeneous, while retaining the spatial layout
23. Application Ⅱ. Image Simplification
• Image simplification
• To prevent over-segmentation, some line segments are removed
• Semantic score of a line segment
𝑆 = 𝑆sl + 𝑆sz + 𝑆ct
• 𝑆sl : the soft-max probability of the line segment
• 𝑆sz : 𝑅 ∪ 𝑄 /(the size of the test image)
• 𝑆ct : 𝜒2 𝐟 𝑅, 𝐟 𝑅∪𝑄 + 𝜒2 𝐟 𝑄, 𝐟 𝑅∪𝑄
𝑅
𝑄
26. Conclusions
• Semantic line detector
• Multi-scale feature map for an image
• Line pooling layer to extract line descriptors
• Multi-task learning for exquisite semantic line detection
• Two applications
• Composition enhancement
• Image simplification