Blind face restoration dataset for Asians
High fidelity facial image datasets for Blind Face Restoration (BFR) models often lack diversity in the distribution of ethnic groups. As a result, BFR models trained on such datasets often produces inaccurate facial structures distinct for less represented ethnic groups. In this report, we delve in...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/175353 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | High fidelity facial image datasets for Blind Face Restoration (BFR) models often lack diversity in the distribution of ethnic groups. As a result, BFR models trained on such datasets often produces inaccurate facial structures distinct for less represented ethnic groups. In this report, we delve into the exploration of enriching facial datasets for BFR models by addressing the inadequacy of Asian facial datasets. To achieve this, we develop a highly automatic and scalable pipeline to collect high quality facial video datasets.
The dataset mirrors the distribution of ethnic groups in Asia, reliably covering the facial image data from diverse ethnic groups in Asia. Additionally, the dataset includes facial images from a range of scenarios and positions from interviews, talk shows and music videos, contributing to improved expression of BFR models. |
|---|