alexlabs / libfacedetection

A binary library for face detection in images. You can use it free of charge with any purpose.

libfacedetection

This is a binary library for face detection in images. The 32-bit and 64-bit dll files are provided.

examples/libfacedetect-example.cpp shows how to use the library.

Comparison on Windows

Method	Time	FPS	Misc
OpenCV	21.2ms	47.2	Yaw angle: -40 to 40 degrees. Classifier: haarcascade_frontalface_alt.xml
frontal	2.7ms	370.4	Yaw angle: -60 to 60 degrees
frontal-surveillance	3.8ms	263.2	Yaw angle: -70 to 70 degrees
multiview	6.6ms	151.5	Yaw angle: -90 to 90 degrees
multiview_reinforce	9.5ms	105.3	Yaw angle: -90 to 90 degrees

• 640x480 image size (VGA), scale=1.2, minimal window size = 48
• Intel(R) Core(TM) i7-4770 CPU @ 3.4GHz
• Multi-core parallelization is enabled for these five methods.

Comparison on iPhone(+)

Method	Time (iPhoneSE)	FPS (iPhoneSE)	Time (iPhone5S)	FPS (iPhone5S)	Misc
frontal	14.9ms	67.1	23.8	42.0	Yaw angle: -60 to 60 degrees
multiview	47.3ms	21.1	75.1	13.3	Yaw angle: -90 to 90 degrees
multiview_reinforce	83.2ms	12.0	132.4	7.6	Yaw angle: -90 to 90 degrees

• 640x480 image size (VGA), scale=1.2, minimal window size = 48
• Multi-core parallelization is disabled.
• C programming language, and no SIMD instruction is used.
• (+) The performance data on iPhone was obtained using a previous version. Currrent version can gain ~40% speed improvement.

Comparison on ARM

Method	Time	FPS	Misc
frontal	21.5ms	46.5	Yaw angle: -60 to 60 degrees
frontal-surveillance	36.5ms	27.4	Yaw angle: -70 to 70 degrees
multiview	67.1ms	14.9	Yaw angle: -90 to 90 degrees
multiview_reinforce	112.6ms	8.9	Yaw angle: -90 to 90 degrees

• 640x480 image size (VGA), scale=1.2, minimal window size = 48
• NVIDIA TK1 "4-Plus-1" 2.32GHz ARM quad-core Cortex-A15 CPU
• Multi-core parallelization is disabled.
• C programming language, and no SIMD instruction is used.

The dll cannot run on ARM. The library should be recompiled from source code for ARM compatibility. If you need the source code, a commercial license is needed.

Evaluation

FDDB: http://vis-www.cs.umass.edu/fddb/index.html

• scale=1.08
• minimal window size = 16
• the heights of the face rectangles are scaled to 1.2 to fit the ground truth data in FDDB.

Tip

• Do NOT use the functions in multiple threads. The memory used in the functions is not protected according to multiple threads.

Author

• Shiqi Yu [e-mail] Computer Vision Institute, Shenzhen University, China

Contributors

• Shengyin Wu, Shenzhen University, China
• Dong Xu, Shenzhen University, China