Diversified Character Dataset for Creative Applications (DCDCA)
Description
Generative Adversarial Networks (GANs) have been widely adapted into several domains. The Implementations of such models demand a large dataset for training, testing, and validation of generated outcomes. Character designers are continuously challenged with creating new concepts for games, videos, and animations; but other than being novel, innovating a new concept of a character must also fit a specific narrative or context. Therefore, despite the gallant strides of success in a wide range of GANs implementations, creating a realistic result of life-like visual scenes and characters does not work well for this creative domain. This dataset is proposed as a cognitive substance that interacts with designers’ competencies in a versatile manner to influence the creative processes of conceptualizing novel characters. Instead of starting from an empty page, designers are given a silhouette of a character manipulated according to the perceptual proposition of designers who interrogate the form as it morphs into a new concept. Designers are also given the option to have a colored/textured version of the silhouette as a possible insight into possible outcomes. The labeled dataset contains three main classes (Men, Women, and Monster) with a total of 22000 characters organized as follows: - 10000 - 512 x 515 Black and white silhouettes of characters collected and processed from different multimedia projects including but not limited to videos, games, and animations. - 6000 – 512 x 512 Black and white silhouettes of characters generated using GANs. - 6000 – 512 x 512 Colored and textured generated Characters using GANs. The three folders are compressed into one zipped file named with the title of the dataset. There is another folder containing approximately 800k Black and white silhouettes that were generated using five different GANs models at different resolutions proposed to assist in calculating Fréchet Inception Distance (FID). The file size of this group is large upon extraction, which is why there were placed in a different directory to download when necessary.
Files
Steps to reproduce
Download and extract: Diversified Character Dataset for Creative Applications.zip The FID calculation dataset (approx. 800k of b/w silhouettes) is placed in a separate directory to be downloaded only when/if needed.