Researchers have introduced FaceMoE, a novel Mixture of Experts (MoE) transformer architecture designed to improve low-resolution face recognition. This architecture employs specialized feed-forward network experts and a top-k router to dynamically assign tokens, promoting expert specialization for different facial regions. FaceMoE aims to enhance feature extraction and aggregation in low-resolution images while mitigating the domain gap between high-resolution and low-resolution data. The model is trained with a combined loss function to ensure expert specialization and stable training, and extensive experiments show it outperforms existing state-of-the-art methods across various benchmarks. AI
IMPACT Introduces a novel MoE architecture for face recognition, potentially improving performance on low-resolution and degraded images.
RANK_REASON The item is a research paper detailing a new model architecture for a specific computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- FaceMoE
- feed forward network
- Gotit.pub
- Hugging Face
- mixture of experts
- ScienceCast
- top-k router
- transformer
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
