← All projects

WACV Oral 2025

PETALface

Adapt a strong face-recognition model to degraded images by training small quality-dependent adapters instead of the entire backbone.

Co-author

Gallery and probe

Original BRIAR examples and image-quality analysis.

How it works

Quality-weighted adapters

PETALface keeps the recognition backbone frozen and trains two LoRA branches, using only 0.48% of the parameters. An image-quality estimate weights their contributions for each input.

Adapting across image quality

Quality-dependent weighting addresses the mismatch between clear gallery images and degraded probes while retaining the pretrained representation. The reported BRIAR comparison uses Swin-B, WebFace4M and Protocol 3.1.

Quality-weighted adapters

Two LoRA branches adapt a frozen backbone using quality-dependent weighting.

Results

The adapters improve face verification compared with the pretrained model and full fine-tuning in this test. TAR is the proportion of genuine matches accepted at a fixed false-accept rate.

Adaptation BRIAR TAR at FAR 0.1%
Pretrained 45.93%
Full fine-tuning 29.68%
PETALface 55.35%

WACV 2025, Table 2. Swin-B, WebFace4M, BRIAR Protocol 3.1.

Higher TAR is better. Only 0.48% of parameters are trained by PETALface.

The comparison uses BRIAR Protocol 3.1 and differs from the FG 2025 BRIAR study. Full fine-tuning reaches 29.68% under this protocol.

Presentation

PETALface gallery/probe input examples and measured recognition results. The photographs are research inputs.