Face detection that fits in a megabyte

What it does Ultra-Light-Fast-Generic-Face-Detector-1MB is a face detection network designed for edge computing devices. The default FP32 PyTorch model is roughly 1.04–1.1 MB, and an INT8-quantized version shrinks to about 300 KB. It comes in two flavors: version-slim, which trims the backbone for speed, and version-RFB, which uses a modified Receptive Field Block to squeeze out extra accuracy.

The interesting bit The repo ships ready-made export formats—ONNX, Caffe, NCNN, and MNN—alongside C++ and Python inference code for ARM and RISC-V runtimes. WIDER FACE benchmarks show the RFB variant outperforming libfacedetection and Retinaface-Mobilenet-0.25 on the Hard set at 320×240, while the slim variant clocks roughly 9.5 ms on a quad-core Raspberry Pi 4B.

Key highlights

• Two variants: version-slim (faster, smaller backbone) and version-RFB (higher accuracy).
• FP32 model size of ~1.04–1.1 MB; INT8 quantized models around ~300 KB.
• About 90–109 MFlops for 320×240 inputs.
• Pre-converted models and inference examples for NCNN, MNN, Caffe, and OpenCV DNN.
• Benchmarked on Raspberry Pi 4B, iPhone 6s Plus, and Kendryte K210.

Caveats

• The README itself warns that benchmark results “may be partially inconsistent.”
• The listed training environment is aging: Python 3.6, PyTorch 1.2, and CUDA 10.0.
• The TODO list still includes adding test data, so the repo feels slightly unfinished.

Verdict Grab it if you need a sub-megabyte face detector for an embedded ARM or RISC-V board and can live with trade-offs in accuracy. Look elsewhere if you need top-tier precision on high-resolution images or a modern training stack.