Reproduce “FaceLiVT: Face Recognition using Linear Vision Transformer with Structural Reparameterization For Mobile Device” 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
Novendra Setyawan, Chi-Chia Sun, Mao-Hsiu Hsu, Wen-Kai Kuo, Jun-Wei Hsieh
FaceLiVT introduces a lightweight hybrid CNN-Transformer model for face recognition, achieving an exceptional balance of high accuracy and low inference latency specifically for mobile devices. It demonstrates significantly faster performance compared to existing state-of-the-art models while maintaining competitive recognition rates across various benchmarks.
Key insights
- Multi-Head Linear Attention (MHLA) effectively reduces the computational complexity of transformer-like layers to linear, enabling efficient global context modeling suitable for mobile devices.
- Structural reparameterization techniques (fused Batch Normalization, reparameterized depthwise convolutions) significantly streamline the inference graph, drastically improving latency on edge hardware without compromising accuracy.
- A holistic approach combining efficient linear attention with inference-time optimizations for CNN components yields a superior accuracy-latency trade-off for mobile face recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.