Reproduce “MobiFace: A Lightweight Deep Learning Face Recognition on Mobile Devices” 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
Chi Nhan Duong, Kha Gia Quach, Ibsa Jalata, Ngan Le, Khoa Luu
MobiFace introduces a lightweight deep neural network architecture designed for face recognition on mobile devices. It achieves 99.72% accuracy on LFW with a 9.3 MB model and 26 ms inference time, and 90.9% on MegaFace, demonstrating performance comparable to larger models at a significantly reduced computational footprint.
Key insights
- Aggressive early downsampling, termed 'Fast Downsampling', in CNNs can significantly reduce computational cost while maintaining robust feature representation for face recognition.
- The use of Bottleneck Residual Blocks with an expansion factor of 2 provides an effective balance between model compactness and the ability to learn rich features.
- Employing a Fully Connected layer for final feature embedding, instead of Global Average Pooling, can yield higher discriminative power by allowing learned weighting of feature map units, leading to improved face recognition accuracy.
- Predicting deep features of a mirrored image from the original image's deep features via a small auxiliary network (Flipped-MobiFace) efficiently enhances pose robustness without requiring a second full network inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.