Supplementary Materials
>Classification results:
Following the same scheme as in [1], we also apply our regression models to a classification task, where we distinguish between high and low quality images. Each test image is assigned a groundtruth label, either high or low quality, by comparing its aesthetic score to two thresholds: 5 − δ/2, 5 + δ/2. The images with aesthetic scores in between the thresholds are discarded.
By increasing δ, we effectively eliminate the ambiguous images, thus increasing the "aesthetics gap" between the two classes and simplifying the classification task. The predicted labels are created by thresholding at the score of 5 [1]. The accuracies are shown below. As expected, with increasing δ, the classification accuracy increases. Similar to the trends in the regression task, our model without sample weights ( NSWR ) achieves the highest classification accuracy on the RS-test while our model with sample weights ( SWR ) further outperforms NSWR on the ED-test .
Classification accuracy (%) of different methods
| |
δ |
0 |
0.1 |
0.5 |
1.0 |
1.5 |
2.0 |
| RS-test |
Lu et al.[2] |
74.46 |
NA |
NA |
NA |
NA |
NA |
| Kao et al.[1] |
71.42 |
72.26 |
76.92 |
82.21 |
85.49 |
89.31 |
| NSWR |
75.73 |
76.91 |
82.20 |
86.84 |
90.82 |
94.12 |
| SWR |
72.4 |
73.22 |
77.66 |
82.28 |
85.41 |
89.87 |
| ED-test |
NSWR |
82.73 |
83.18 |
85.11 |
86.62 |
87.21 |
86.95 |
| SWR |
83.46 |
83.95 |
86.09 |
88.24 |
89.62 |
89.96 |
>References:
[1] Kao, Y., Wang, C. and Huang, K., Visual aesthetic quality assessment with a regression model. In ICIP, 2015.
[2] Lu, X., Lin, Z., Jin, H., Yang, J. and Wang, J.Z., Rapid: Rating pictorial aesthetics using deep learning. In ACMMM , 2014.
>Results of the image enhancement application:
We apply our regression CNN model with sample weights ( SWR ) to an image enhancement application, where we obtain aesthetically appealing crops of the input images.
We only show 50 sample results here due to the size limitation of the supplementary materials.
These images are not part of the AVA dataset but random collections from the Internet.
The results show that our model, while only trained on the AVA dataset, manages to assess aesthetics for random collected images from the Internet.
Note that images are fitted to the cells, hence the size of the images seen in the table do not represent the actual size. Click on the image to check the actual size of the image.
