Evaluation Metrics for Line Art Colorization
Year: 2023
Type: dissertation
Abstract: Line art colorization is a necessary and time-consuming process in creating 2D illustrations, colorized mange and animation.It is extremely time-consuming and labor intensive because every single step of the colorization has to be done by hand, and it contains a lot of redundant work.Although progress has been done in automating this process, currently it is still dominantly done manually given the inconsistency and low quality of automatic approaches.One of the main reasons is the lack of proper evaluation, there are no evaluation method designed specifically for the evaluation of line art colorization results, most of the line art colorization research solely relied on simply using the Fréchet Inception Distance (FID) for evaluation, and ignored the correlation between the hints and the performance.In this thesis, we proposed an evaluation method specifically for the evaluation of line art colorization, aiming to explore the correlation between the hints and performance, namely the number of hints and the type of hints.Furthermore, we collect and process a large scale dataset containing sketch and color image pairs that is suitable for the training of line art colorization.We also designed an algorithm that can generate different types of hints deterministically, namely controlling the location, size and number of hints generated, which can produce proper evaluation dataset for line art colorization, which is crucial in evaluating the performance between different models.
Authors: Yuan, Mingcheng, 133133
Institution: Waseda University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Keyword: evaluation metrics
Open Access status: green
OpenAlex ID: W6981640593
