Week 7 — Image Transformation According to Art Style
Hi everyone! Last week, we have talked about the results of our style-based clustering and object-based clustering methods. Today, we are going to talk about our final phase of the project.
After running the model on different clusters, we run the model on different clusters. But, some of the clusters didn’t give accurate results. Object-based clusters haven’t good results. So, we focused on style-based clustered images.
Evaluation
In the Pros and Cons of the GAN Evaluation Measures paper, 24 different evaluation methods are mentioned. These are separated into 2 different quantitative and qualitative subtitles. According to the quantitative, results are evaluated by more objective methods based on numerical scores. Some of the methods include pre-trained models to classify generated results like Inception Score and Frechet Inception Distance. However, we can’t use these methods because they classify more object-based results.
Qualitative evaluation metrics are mostly human-based evaluation techniques. Preference Judgement is one of these techniques. According to these techniques, results are mostly evaluated by a participation group. But here, the participants should have the necessary information about the transformed domain. And, we firstly worked on more participants in our close friends. However, after our lecturer’s feedback, we noticed that more domain expert people should take place in the participant group. So, we prepare one extra user story after that.
We use 10 different Van Gogh style transformed image to ask 6 art expert from art department students. We asked them “Does these images look like one of Van Gogh’s styles?”. According to their rates, the model trained over the style-cluster 1 dataset gave better results. If you remember from last week, we got 3 different style based clusters and cluster 1 gave us more the first thing coming to mind when Van Gogh’s art is considered. Let’s we talked about our model’s strengths and weaknesses.
Strengths
- If the artist’s paintings contain many pictures and have different styles, we divide the data set into clusters to obtain better-styled pictures.
- When the artworks are divided into clusters, we can get better results on the specific paintings of an artist.
- While the number of artist’s artworks is high, our model saves time.
Weaknesses
- Our model cannot be trained very much with a small dataset.
- If the artist’s artworks have a few numbers, of images, our model cannot be trained very efficiently.
- If the artist has only one style, clustering cannot be done.
- If the artist has too many styles, the dataset is divided into too many clusters.
Future Works
If we consider how our model can be improved and what high-level methods can be used:
- Graph-Based Representativity Learning methods can be used to choose the presentative paintings.
- We can try to extract style-based features with more complex feature extraction methods. For example, RestNet50.
- We can examine the effects of clustered train sets on other generative models such as GAN, cGAN wGAN.
Contributors:
References:
- https://arxiv.org/pdf/1508.06576.pdf, Pros and Cons of GAN Evaluation Measures
