Reproduce “ViTNT-FIQA: Training-Free Face Image Quality Assessment with Vision Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guray Ozgur, Eduarda Caldeira, Tahar Chettaoui, Jan Niklas Kolf, Marco Huber, Naser Damer, Fadi Boutros
ViTNT-FIQA is introduced as a training-free face image quality assessment method that leverages the stability of patch embedding evolution across intermediate Vision Transformer (ViT) blocks. The approach achieves competitive performance with state-of-the-art FIQA methods on various benchmarks, requiring only a single forward pass without additional training or backpropagation.
Key insights
- High-quality face images exhibit stable and smooth feature refinement trajectories across consecutive ViT blocks, as evidenced by minimal changes in their patch embeddings.
- The stability of intermediate patch embedding evolution within pre-trained ViT models serves as a novel, intrinsic, and computationally efficient indicator of face image quality.
- ViT's learned self-attention weights can effectively identify and assign importance to facial regions crucial for quality assessment, improving the overall image-level quality score aggregation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.