Video Super-Resolution Project
-Video Demo

1Chih-Chun Hsu, 1Chia-Wen Lin, and Li-Wei Kang

1Department of Electrical Engineering
National Tsing Hua University
Hsinchu 30013, Taiwan

Description

There are six videos in each demo video. We compare the proposed method with sparse coding SR, ASDS SR, NLBP SR, and TSS-SR. The naïve resolutions for video #5, #7, and #8 is 1280x720, while the resolution for the rest videos is 640x480. To have a fair comparison, we suggest that the resolution of YouTube player can be changed to the highest resolution.

All demo videos (without cropping) can be downloaded from Here.

Menu

 

Video Demo (Fine and dynamic textures)

Video #1: Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
We can see that the proposed method can achieve fine-detail reconstruction with temporal coherent result. Although the reconstructed video using TS-SR achieves fine-detail result, the temporal incoherency problem is presented. NLBP shows the over-sharpness property, ASDS and Sc-SR show the over-smooth results.

 

Table #1 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #1

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 80.00% 85.00% 90.00% 95.00% 90.00% 88.00%
ASDS- SR  20.00% - 50.00% 60.00% 90.00% 70.00% 58.00%
NLIBP-SR  15.00% 50.00% - 60.00% 90.00% 70.00% 57.00%
SC-SR 10.00% 40.00% 40.00% - 80.00% 60.00% 46.00%
Bicubic 5.00% 10.00% 10.00% 20.00% - 40.00% 17.00%
TS-SR 10.00% 30.00% 30.00% 40.00% 60.00% - 34.00%



Video #2: Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: ScSR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
In this video, we can see that the reconstructed video of our method provides more fine-detail result than that of other methods. The dynamic background (grass) also keeps temporal coherency.

Table #2 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #2

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 75.00% 80.00% 85.00% 90.00% 85.00% 83.00%
ASDS- SR  25.00% - 60.00% 60.00% 90.00% 70.00% 61.00%
NLIBP-SR  20.00% 40.00% - 55.00% 90.00% 75.00% 56.00%
SC-SR 15.00% 40.00% 45.00% - 90.00% 60.00% 50.00%
Bicubic 10.00% 10.00% 10.00% 10.00% - 40.00% 16.00%
TS-SR 15.00% 30.00% 25.00% 40.00% 60.00% - 34.00%

 


 

Video #3: Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
The rendered & SR video using the proposed method shows high-quality reconstruction with temporal coherency. Compared to other SR methods, the proposed method presents superior performance.

Table #3 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #3

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 85.00% 80.00% 85.00% 90.00% 85.00% 85.00%
ASDS- SR  15.00% - 55.00% 55.00% 80.00% 60.00% 53.00%
NLIBP-SR  20.00% 45.00% - 60.00% 70.00% 75.00% 54.00%
SC-SR 15.00% 45.00% 40.00% - 70.00% 55.00% 45.00%
Bicubic 10.00% 20.00% 30.00% 30.00% - 45.00% 27.00%
TS-SR 15.00% 40.00% 25.00% 45.00% 55.00% - 36.00%


 

Video #4: Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
In this video, the fast motion is presented. We can observed that the proposed method still generates high-quality reconstruction, which outperforms the results of other methods. Besides, the dynamic background (grass) can be well reconstructed in a temporal coherent way, while the still-background (i.e. stonewalling) is also reconstructed without jitter-like artifacts, compared to TS-SR.

Table #4 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #4

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 85.00% 90.00% 90.00% 90.00% 95.00% 90.00%
ASDS- SR  15.00% - 60.00% 70.00% 80.00% 65.00% 58.00%
NLIBP-SR  10.00% 40.00% - 60.00% 80.00% 65.00% 51.00%
SC-SR 10.00% 30.00% 40.00% - 95.00% 50.00% 45.00%
Bicubic 10.00% 20.00% 20.00% 5.00% - 40.00% 19.00%
TS-SR 5.00% 35.00% 35.00% 50.00% 60.00% - 37.00%


Video Demo (Mixed textures and non-textures)

 

Video #5 (Lucy): Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
It is clear that the details of building, especially in window, significantly outperform other methods like ASDS, NLIBP. It is remarkable that the performance of the proposed method for general video is comparable with other advanced SR methods.

Table #5 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #5

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 80.00% 90.00% 90.00% 95.00% 90.00% 89.00%
ASDS- SR  20.00% - 45.00% 65.00% 90.00% 70.00% 58.00%
NLIBP-SR  10.00% 55.00% - 60.00% 80.00% 70.00% 55.00%
SC-SR 10.00% 35.00% 40.00% - 80.00% 55.00% 44.00%
Bicubic 5.00% 10.00% 20.00% 20.00% - 40.00% 19.00%
TS-SR 10.00% 30.00% 30.00% 45.00% 60.00% - 35.00%


 

Video #7 (Campus): Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
The larger motion of the dynamic textures (tree) result in smoothing synthesized dynamic textures. However, we still want to point out that the visual quality of the reconstructed video using the proposed method still slightly outperforms that of other methods.

Table #6 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #7

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 65.00% 75.00% 80.00% 90.00% 90.00% 80.00%
ASDS- SR  35.00% - 65.00% 60.00% 85.00% 75.00% 64.00%
NLIBP-SR  25.00% 35.00% - 50.00% 85.00% 85.00% 56.00%
SC-SR 20.00% 40.00% 50.00% - 90.00% 60.00% 52.00%
Bicubic 10.00% 15.00% 15.00% 10.00% - 55.00% 21.00%
TS-SR 10.00% 25.00% 15.00% 40.00% 45.00% - 27.00%


 

Video #8 (Ocean): Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
In this video, the dynamic textures is quite limited, but synthesized result still outperforms other methods.

Table #7 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #8

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 85.00% 80.00% 95.00% 95.00% 85.00% 88.00%
ASDS- SR  15.00% - 65.00% 75.00% 90.00% 70.00% 63.00%
NLIBP-SR  20.00% 35.00% - 75.00% 85.00% 75.00% 58.00%
SC-SR 5.00% 25.00% 25.00% - 90.00% 45.00% 38.00%
Bicubic 5.00% 10.00% 15.00% 10.00% - 40.00% 16.00%
TS-SR 15.00% 30.00% 25.00% 55.00% 60.00% - 37.00%


 

Video #6 (Building): Top-left: Ground truth, Top-middle: NLIBP-SR [6], Top-right: The proposed method, Bottom-left: Sc-SR [9], Bottom-middle: ASDS-SR [11], and Bottom-right: TS-SR [26].
In this video, the details of the building is well reconstructed using TS-SR. The proposed DTS-SR method is further used to maintain the temporal consistence. In conclusion, the proposed method has superior performance over existing SR methods.

Table #8 Subjective “visual quality” evaluation by paired comparisons (in relative winning percentage) for the HR video #6

Method Proposed ASDS- SR  NLIBP-SR  SC-SR  Bicubic  TS-SR  Average 
Proposed - 90.00% 85.00% 95.00% 95.00% 85.00% 90.00%
ASDS- SR  10.00% - 55.00% 60.00% 90.00% 70.00% 57.00%
NLIBP-SR  15.00% 45.00% - 75.00% 85.00% 75.00% 59.00%
SC-SR 5.00% 40.00% 25.00% - 90.00% 60.00% 44.00%
Bicubic 5.00% 10.00% 15.00% 10.00% - 50.00% 18.00%
TS-SR 15.00% 30.00% 25.00% 40.00% 50.00% - 32.00%


 

References

[6] W. Dong, L. Zhang, G. Shi, and X. Wu, “Nonlocal back-projection for adaptive image enlargement,” in Proc. IEEE Int. Conf. Image Process., Cairo, Egypt, Nov. 2009, pp. 349352.
[9] J. Yang, J. Wright, T. Huang, and Y. Ma, “Image super-resolution via sparse representation,” IEEE Trans. Image Process., vol. 19, no. 11, pp. 2861–2873, Nov. 2010.
[11] W. Dong, L. Zhang, G. Shi, and X. Wu, “Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization,” IEEE Trans. Image Process., vol. 20, no. 7, pp. 18381857, July 2011.
[26] Y. HaCohen, R. Fattal, and D. Lischinski, “Image upsampling via texture hallucination,” in Proc. IEEE Int. Conf. Comput. Photography, Cambridge, MA, USA, pp. 2030, Mar. 2010.