ViT-HarDis improves face recognition by conditioning attention on sample difficulty

Researchers have developed ViT-HarDis, a novel method to improve face recognition in Vision Transformers by conditioning attention head disentanglement on sample difficulty. This approach applies stronger decorrelation to harder samples, where identity ambiguity is highest, without introducing architectural changes or increasing inference costs. Experiments demonstrate that ViT-HarDis achieves state-of-the-art results on several benchmarks, including IJB-B and IJB-C, by reducing the similarity of hard samples to competing class centers and decreasing inter-head attention similarity. AI

IMPACT Enhances face recognition accuracy in Vision Transformers by optimizing attention mechanisms for challenging samples.

RANK_REASON The cluster contains a research paper detailing a new method for improving Vision Transformer performance on face recognition tasks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• IEEE/CVF Winter Conference on Applications of Computer Vision
• IJB-B
• IJB-C
• MS1MV3
• Tahar Chettaoui
• TinyFace
• Vision Transformers
• ViT-HarDis
• WACV 2027
• WebFace4M

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →