SegFace - Face Segmentation of Long-Tail Classes
Abstract
- Face Parsing -> Semantic Segmentation
- CNN-based methods => head classes perform well, but long-tail classes perform poorly
- Transformers => long-tail classes perform better, F1 score also improves and edge device performance is better
Introduction
Face Parsing Purpose
- face swapping
- face editing
- face generation
- face completion
- facial makeup
What is Long-Tail Classes?
In Face Long-Tail classes like accessoris: eyeglasses,earring, hats,necklaces
Problem of Long-Tail Classes
- lower probability of appearance in dataset
- less training data
- small item annotation, spend more time and money
Method of model
- AGRNet
- EAGRNet
- FaRL
- ImageNet
- DML-CSR
- FP-LIIF