Paper Presentation – CG Fall 2019

Slide 1

Published in: 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

​

​

​

​

​

Presenter: Le Ngoc Hanh

Slide 2

Outline

1. Introduction
2. Geometric change estimation
3. Aspect Ratio Similarity (ARS)
4. Conclusion

​

Slide 3

Outline

1. Introduction
2. Geometric change estimation
3. Aspect Ratio Similarity (ARS)
4. Conclusion

​

Slide 4

1. Introduction

Slide 5

Original image

Size: H x W

Retargeted image

Size: h’ x w’

CAIR^* operator

(*): Content Aware Image Retargeting

1. Introduction

Slide 6

From its beginning, traditional resizing methods:

(*)

(*)

(*): source from wikipedia

1. Introduction

Slide 7

• Continuous methods:
• Add a deformed mesh 🡪 then resize the mesh under constraint equations.
• [Scaling, Scale-and-Stretch, Streaming Video, Warping]

• Discrete methods:
• Treat images as pixels 🡪 the deformation is mainly due to the retention or discarding of content.
• [Seam Carving, Cropping, Shift-Map]

CAIR operators

1. Introduction

Slide 8

Original image

Size: H x W

Retargeted image

Size: h’ x w’

CAIR^* operator

(*): Content Aware Image Retargeting

Quality for the images after retargeting!!!

1. Introduction

• Some previous works related to measuring the quality of the Image Retargeting.

Slide 9

1. Introdution

Slide 10

1. Introduction

Slide 11

1. Introduction

Slide 12

1. Introduction

Slide 13

1. Introduction

• Major drawback of these works:
• The removed or squeezed image content in original image are allowed to be matched with pixles in the retargeted image .

🡪 they are not necessary and inhibit the estimation of real geometric change.

Slide 14

Outline

1. Introduction
2. Geometric change estimation
3. Aspect Ratio Similarity (ARS)
4. Conclusion

​

Slide 15

2. Geometric change estimation

• Matching energy function: consider about the data term and smoothness term.

​

Slide 16

Find resampling location for each pixel

Slide 17

Base on the geometric change, we can:

• Estimate the cropping window ideally for CR;
• Find out the uniform squeeze for SCL;
• Effectively recover the removed seams for SC;
• Estimate the seriously warped region for WARP.

​

Slide 18

2. Geometric change estimation

• Almost all the quality degradation in the retargeted image is related to the geometric change .
• Based what we estimate, the visual quality of the retargeted image can be effectively evaluated ��

​

Slide 19

Base on the geometric change, we can:

• Estimate the cropping window ideally for CR;
• Find out the uniform squeeze for SCL;
• Effectively recover the removed seams for SC;
• Estimate the seriously warped region for WARP.

2. Geometric change estimation

Slide 20

Geometric change is the major reason for quality degradation of the retargeted image (IR).

🡪 Propose a measurement of quality of IR based on the estimation and evaluation of the geometric change.

Outline

1. Introduction
2. Geometric change estimation
3. Aspect Ratio Similarity (ARS)
4. Conclusion

​

Slide 21

3. Aspect Ratio Similarity (ARS)

• Establish the local geometric change from (O) to (R) by dividing (O) and (R) into blocks. The ARS of each block pair indicate its information loss and visual distortion.

Slide 22

Original image (O)

Retargeted image(R)

Estimate locations for all the pixels in (R)

3. Aspect Ratio Similarity (ARS)

Slide 23

Define the saliency map 🡪 to obtain the perceptual quality for the whole image

3. Aspect Ratio Similarity (ARS)

Slide 24

close or equal to zero, it indicates that the retargeted block is suffering from serious information loss and distortion or even removed totally .

close to 1, the block content in original image is generally kept in high quality in retargeted image.

Outline

1. Introduction
2. Geometric change estimation
3. Aspect Ratio Similarity (ARS)
4. Conclusion

​

Slide 25

4. Conclusion

Slide 26

4. Conclusion

• Contributions:
• All the quality degradation in the retargeted image is related to the geometric change 🡪 this paper propose an ARS metric on the estimation and evaluation of the geometric change during image retargeting.
• Compared to other state-of-the-art methods in the widely used dataset, the ARS metric yields statistically better results in the prediction accuracy.

�

Slide 27

• Drawback:
• ARS does not detect the overall structure change of the image 🡪 this evalution is not comprehensive.
• What if the saliency map method employed in this research is not good?

��

End.

Slide 28

Thanks for your listening