Human Image Transfer: From Natural Settings to Controlled Studios
Ladda ner
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Complex adaptive systems (MPCAS), MSc
Publicerad
2023
Författare
NORDIN, ISAC
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Machine learning (ML) and artificial intelligence (AI) are two concepts that have
driven advancements in various fields, including the fashion industry. Virtual try on, the technology which allows a user to change a person’s clothes in an image,
has gained more attention. However, their potential cannot be fully utilized when
applied to photos captured in real-world environments (in-the-wild). Therefore, this
project aims to investigate methods for realistically transferring in-the-wild photos
to in-studio photos, focusing on domain transfer models.
To do this, four CNN-based models were used, trained to fix augmented in-studio images to their original state. The augmentations attempted to simulate in-the-wild
images by swapping the background and making the image brighter and blurrier
among others. To help the model discern the background of the person a segmentation extractor was used and evaluated. To get the final model, various experiments
were done. The model demonstrated its ability to remove lighting, fix sharpened
images, and remove noise, but failed at removing shadows among other things. The
model showcased better performance at transferring in-the-wild images to in-studio
images than copy-pasting the person into a studio background. The segmentation
played an important role, ignoring to include body parts that were not inside the
segmentation. The evaluation method showcased inconsistencies and needs further
research.
Beskrivning
Ämne/nyckelord
Deep learning, Domain transfer, Virtual Try on