lvpchen / opennpd

C++ detect and train of "A Fast and Accurate Unconstrained Face Detector".

OpenNPD

Project of object detection.

Usage:

1. Add needed librarys to 3rdpart folder(opencv).

2. Make: cd src; make clean; make RELEASE=1; cd ..;

3. Detect example: sh ./script/test_detect_demo.sh

4. Training example: Please refer to ./script/npd_train_demo.sh

Note:

1. Thereis a npddetect::prescandetect function for faster detection with some lose on recall. Additional parameter stepR refers to pre-scan step size compared to the original scan step size ( float more than 1 ) . The thresR refers to the threshold to reject the window( float in [0-1] ).

Result:

• ROC:

  

• Speed:

| image size | window size | cores | time (ms) | | :---: | :---: | :---: | :---: | :---: | | 640x480 | 20x20 | 1 | ~50 |

• ROC for npddetect::prescandetect:

• Speed for npddetect::prescandetect:

params	image size	window size	time(ms)
none	1920x1080	20x20	532.400239
stepR = 2, thresR = 0.2	1920x1080	20x20	344.154205
stepR = 2, thresR = 0.3	1920x1080	20x20	282.128798
stepR = 3, thresR = 0.2	1920x1080	20x20	286.230415
stepR = 3, thresR = 0.3	1920x1080	20x20	226.091203
stepR = 4, thresR = 0.3	1920x1080	20x20	202.147923

#License and Citation

This software is free for noncommercial use. This software is provided "as is", without any warranty of upgradation or customized development. It is your own risk of using this software. The authors are not responsible for any damage caused by using this software.

References:

This software is based on the MATLAB edition. Thanks for the work of Liao et al. Project page.

@article{

 Author = {Shengcai Liao, Member, IEEE, Anil K. Jain, Fellow, IEEE, and Stan Z. Li, Fellow, IEEE},

 Title = {A Fast and Accurate Unconstrained Face Detector},

 Year = {2014}

}