Master thesis presentation Isac Nordin, MPCAS

Abstract: Machine learning (ML) and artificial intelligence (AI) have driven advancements in various fields, including the fashion industry. Virtual try-on, the ability to change a person’s clothes in an image, has gained popularity, but it faces challenges when used on individuals captured in real-world environments (in-the-wild). This project aims to investigate methods for realistically transferring in-the-wild photos to in-studio photos, focusing on domain transfer models. The discussion highlights the evaluation metrics used in assessing the performance of the models. While certain metrics like PSNR show high scores, they do not reflect the overall quality of the generated images. The importance of feature extractors, particularly segmentation, is emphasized, as it heavily influences the quality of the generated images. Improvements suggested include in-painting the human onto a general background and using an Image Matting extractor for accurate transfer of hair and glasses. The models’ performance is discussed, indicating that the SEG- GAN model performs slightly better overall. However, the effects achieved, such as lighting changes and removal of shadows, are minimal. Augmentations can be explored to address issues like blurriness, shadows, and lighting, but further experimentation is needed.

Password: 257914

Supervisor: Fredrik Viberg
Examiner: Giovanni Volpe
Opponent: Kelly Ma