ShiqiYu/libfacedetection

libfacedetection

This is an open source library for CNN-based face detection in images. The CNN model has been converted to static variables in C source files. The source code does not depend on any other libraries. What you need is just a C++ compiler. You can compile the source code under Windows, Linux, ARM and any platform with a C++ compiler.

SIMD instructions are used to speed up the detection. You can enable AVX2 if you use Intel CPU or NEON for ARM.

The model files are provided in src/facedetectcnn-data.cpp (C++ arrays) & the model (ONNX) from OpenCV Zoo. You can try our scripts (C++ & Python) in opencv_dnn/ with the ONNX model. View the network architecture here.

Please note that OpenCV DNN does not support the latest version of YuNet with dynamic input shape. Please ensure you have the exact same input shape as the one in the ONNX model to run latest YuNet with OpenCV DNN.

examples/detect-image.cpp and examples/detect-camera.cpp show how to use the library.

The library was trained by libfacedetection.train.

How to use the code

You can copy the files in directory src/ into your project, and compile them as the other files in your project. The source code is written in standard C/C++. It should be compiled at any platform which supports C/C++.

Some tips:

• Please add facedetection_export.h file in the position where you copy your facedetectcnn.h files, add #define FACEDETECTION_EXPORT to facedetection_export.h file. See: issues #222
• Please add -O3 to turn on optimizations when you compile the source code using g++.
• Please choose 'Maximize Speed/-O2' when you compile the source code using Microsoft Visual Studio.
• You can enable OpenMP to speedup. But the best solution is to call the detection function in different threads.

You can also compile the source code to a static or dynamic library, and then use it in your project.

How to compile

CNN-based Face Detection on Intel CPU

Using AVX2 instructions

Using AVX512 instructions

• Minimal face size ~10x10
• Intel(R) Core(TM) i7-7820X CPU @ 3.60GHz
• Multi-thread in 16 threads and 16 processors.

CNN-based Face Detection on ARM Linux (Raspberry Pi 4 B)

• Minimal face size ~10x10
• Raspberry Pi 4 B, Broadcom BCM2835, Cortex-A72 (ARMv8) 64-bit SoC @ 1.5GHz
• Multi-thread in 4 threads and 4 processors.

Performance on WIDER Face

Run on default settings: scales=[1.], confidence_threshold=0.02, floating point:

AP_easy=0.887, AP_medium=0.871, AP_hard=0.768

Highway Optimized Version

An independent Highway-based implementation has been added under highway/. It keeps the original implementation untouched and exposes a separate C API:

#include "facedetect_hw.h"

int* results = facedetect_hw_cnn(result_buffer, bgr_image_data,
                                 width, height, step);

The Highway version follows the same deployment model as the original project: build separately for each instruction set/platform. It does not currently use Highway runtime dynamic dispatch. The current x86 performance path is:

The public API uses thread-local internal workspaces, so the recommended multi-threading model is external parallelism: call facedetect_hw_cnn from multiple threads, with one result buffer per calling thread.

Measured resolution benchmark:

Notes:

• Benchmark command: fdt_hw_resolution_benchmark.exe images\cnnresult.png 128 28.
• Test CPU: Intel(R) Core(TM) i7-14700KF, 20 cores, 28 logical processors, max clock reported by Windows: 3400 MHz.
• Compiler/toolchain: Visual Studio 17 2022, MSVC 19.40.33821.0, Release build.
• origin AVX2 was built with _ENABLE_AVX2; origin default/scalar was built without original SIMD macros.
• Highway scalar was built with FDT_HW_FORCE_SCALAR=ON.
• External MT means external multi-threading with 28 caller threads and one result buffer per thread. The reported time is throughput-normalized average time per image.
• AVX512 builds are buildable with /arch:AVX512, but cannot be run on the test CPU, so no AVX512 timing is reported here.
• ARM Linux has not been tested for the Highway version yet.
• Existing ARM Linux numbers below are from the original implementation, not the new Highway implementation.
• See COMPILE.md and example/detect-image-highway.cpp / example/benchmark-highway.cpp for build and usage examples.

Rust Crate

A pure Rust implementation is available under rust/ as the crate libfacedetection_rs. It uses the same CNN model data, generated into Rust source so the crate is self-contained and does not need to parse the C++ model file at build time.

The Rust crate is intended for normal Rust dependency usage:

[dependencies]
libfacedetection_rs = "0.1"

Minimal usage:

use libfacedetection_rs::{image, Detector};

fn main() -> Result<(), libfacedetection_rs::DetectError> {
    let image = image::open("face.jpg")?;
    let mut detector = Detector::new();
    let detection = detector.detect(&image)?;

    println!("faces: {}", detection.face_count());
    Ok(())
}

Current crate shape:

Measured resolution benchmark against the original C++ implementation:

Notes:

• Benchmark image: images\cnnresult.png, resized to the listed input resolutions before detection.
• C++ command: .\build-hw-bench-avx2-hybrid\Release\fdt_hw_resolution_benchmark.exe images\cnnresult.png 128 28.
• Rust command: cargo run --release --manifest-path rust/dev/resolution_benchmark/Cargo.toml -- images\cnnresult.png 128 28.
• Test CPU: Intel(R) Core(TM) i7-14700KF, 20 cores, 28 logical processors, max clock reported by Windows: 3400 MHz.
• C++ compiler/toolchain: Visual Studio 17 2022, MSVC 19.40.33821.0, Release build. The origin C++ row above is built with _ENABLE_AVX2.
• Rust toolchain: rustc 1.95.0, cargo 1.95.0, release profile. The Rust row uses runtime AVX2 dispatch when available.
• External MT means external multi-thread throughput with 28 caller threads. The C++ benchmark uses one result buffer per caller thread. The Rust benchmark uses one reusable Detector per worker thread, warms each worker, then synchronizes the timed loop.
• The reported MT time is throughput-normalized average time per image, not single-request latency.
• The C++ benchmark uses OpenCV resize; the Rust benchmark uses the image crate resize path. The table is intended for performance comparison, not result-buffer parity or accuracy evaluation.
• See rust/README.md for crate usage and rust/docs/rs-performance-optimization-report.md for the detailed optimization report.

Author

• Shiqi Yu, [email protected]

Contributors

All contributors who contribute at GitHub.com are listed here.

The contributors who were not listed at GitHub.com:

• Jia Wu (吴佳)
• Dong Xu (徐栋)
• Shengyin Wu (伍圣寅)

Acknowledgment

The work was partly supported by the Science Foundation of Shenzhen (Grant No. 20170504160426188).

Citation

The master thesis of Mr. Wei Wu. All details of the algorithm are in the thesis. The thesis can be downloaded at 吴伟硕士毕业论文

@thesis{wu2023thesisyunet,
    author      = {吴伟},
    title       = {面向边缘设备的高精度毫秒级人脸检测技术研究},
    type        = {硕士学位论文},
    institution = {南方科技大学},
    year        = {2023},
}

The paper for the main idea of this repository https://link.springer.com/article/10.1007/s11633-023-1423-y.

@article{wu2023miryunet,
	title     = {YuNet: A Tiny Millisecond-level Face Detector},
	author    = {Wu, Wei and Peng, Hanyang and Yu, Shiqi},
	journal   = {Machine Intelligence Research},
	pages     = {1--10},
	year      = {2023},
	doi       = {10.1007/s11633-023-1423-y},
	publisher = {Springer}
}

The survey paper on face detection to evaluate different methods. It can be open-accessed at https://ieeexplore.ieee.org/document/9580485

@article{feng2022face,
	author  = {Feng, Yuantao and Yu, Shiqi and Peng, Hanyang and Li, Yan-Ran and Zhang, Jianguo},
	journal = {IEEE Transactions on Biometrics, Behavior, and Identity Science}, 
	title   = {Detect Faces Efficiently: A Survey and Evaluations}, 
	year    = {2022},
	volume  = {4},
	number  = {1},
	pages   = {1-18},
	doi     = {10.1109/TBIOM.2021.3120412}
}

The loss used in training is EIoU, a novel extended IoU. The paper can be open-accessed at https://ieeexplore.ieee.org/document/9429909.

@article{peng2021eiou,
	author  = {Peng, Hanyang and Yu, Shiqi},
	journal = {IEEE Transactions on Image Processing}, 
	title   = {A Systematic IoU-Related Method: Beyond Simplified Regression for Better Localization}, 
	year    = {2021},
	volume  = {30},
	pages   = {5032-5044},
	doi     = {10.1109/TIP.2021.3077144}
}