face detection
Ultra-Light-Fast-Generic-Face-Detector-1MB is a compact (~1MB) CNN face detector with PyTorch training code and NCNN/MNN/Caffe/ONNX inference examples for edge devices.
Same problem, same approach. Swapping one for another is a config change, not a rewrite.
Use when you want a current, self-contained C++ face detector with 1000 FPS rather than a PyTorch-trainable model with NCNN/MNN exports.
An open source library for face detection in images. The face detection speed can reach 1000FPS.
Solves the same problem with a different architecture or at a different layer. Expect to rewrite the integration.
Use when you need a Darknet-based face+landmark detector with ncnn support and can tolerate an old, unmaintained codebase.
加入关键点的darknet训练框架，轻量级的人脸检测，支持ncnn推理
Use when you need YOLOv3-family face detection plus landmarks in PyTorch and can accept more compute than the seed.
yoloface大礼包 使用pytroch实现的基于yolov3的轻量级人脸检测（包含关键点）
Use the YoloFace-500k variant here when a 420KB YOLO-style face model in a generic MobileNet-YOLO repo is more important than the seed's dedicated face-detection examples.
MobileNetV2-YoloV3-Nano: 0.5BFlops 3MB HUAWEI P40: 6ms/img, YoloFace-500k:0.1Bflops 420KB:fire::fire::fire:
Use when you need RetinaFace detection in C++ and are willing to accept an old, unmaintained implementation.
RetinaFace detector with C++
Use when you need RetinaFace's higher WiderFace accuracy and are willing to run a heavier PyTorch model than the seed.
Retinaface get 80.99% in widerface hard val using mobilenet0.25.
Use if your stack is TensorFlow and you want a WIDER-trained MobileNet-SSD face detector, accepting a heavier/unmaintained TensorFlow OD API pipeline.
A mobilenet SSD based face detector, powered by tensorflow object detection api, trained by WIDERFACE dataset.
Use when you want to train RetinaFace on custom data in PyTorch and do not need the seed's edge-optimized lightweight deployment.
这是一个retinaface-pytorch的源码，可以用于训练自己的模型。
Use when you prefer this independent PyTorch RetinaFace implementation for custom training and can accept a heavier model.
Reimplement RetinaFace with Pytorch
Use when you need YOLOv7 face boxes plus landmarks and have GPU compute available; it is not an edge-sized 1MB model.
yolov7 face detection with landmark
Use when anchor-free CenterFace face+landmark detection fits your accuracy/speed needs and MNN/NCNN portability is less important.
face detection
Use when you specifically want the LFFD method with MNN inference on edge devices and can tolerate a stale repository.
MNN Implementation of the paper of LFFD: A Light and Fast Face Detector for Edge Devices
Use when you want a maintained pip-installable Python RetinaFace library with a high-level API and do not need the seed's edge-focused 1MB model.
RetinaFace: Deep Face Detection Library for Python
An earlier or more limited way to do the job. Still the right call when you need something small, proven, or CPU-only.
Previous-generation MATLAB/CVPR 2017 method for detecting very tiny faces; choose only for reproducing that research, not for current edge deployment.
Tiny Face Detector, CVPR 2017
Does the same job, but ships as an app. Useful to a person, not swappable into a codebase.
Complete Android SCRFD face-detection sample/demo; useful when you want a ready app to copy rather than a model to integrate into your own codebase.
ncnn android scrfd face detection
These projects were analysed and named Ultra-Light-Fast-Generic-Face-Detector-1MB among their alternatives. The relationship is not symmetric — how Ultra-Light-Fast-Generic-Face-Detector-1MB rates them is a separate judgement, made when Ultra-Light-Fast-Generic-Face-Detector-1MB is analysed in its own right.
calls Ultra-Light-Fast-Generic-Face-Detector-1MB “A lightweight face detector, narrower and older.”
State-of-the-art 2D and 3D Face Analysis Project