ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | ros @ Robotics Stack Exchange
Ask Your Question

Improving face recognition with cob_people_detection?

asked 2016-09-02 13:08:26 -0500

Pi Robot gravatar image


I am using the most excellent cob_people_detection package under ROS Indigo and Ubuntu 14.04 together with an MS Kinect (Xbox 360) at 640x480 resolution. I have only 5 people so far in the database with 40 images per person taken at different angles but the face recognizer often assigns the wrong label to one of those same people standing in front of the camera. I've tried Fischer faces, EigenFaces, and LDA2D but performance does not change much.

Does anyone have any suggestions for improving recognition performance?


edit retag flag offensive close merge delete


I tend to use the people package for this type of work. It uses the opencv haar classifiers, which are pretty good. There's many of them and you can switch them out easily.

curranw gravatar image curranw  ( 2016-09-02 22:28:52 -0500 )edit

Thanks @curranw, but as far as I can tell, the people package only does face detection whereas my issue is with face recognition. The cob_people_detection package actually does a great job at face detection (as well as head detection even when the face is not visible).

Pi Robot gravatar image Pi Robot  ( 2016-09-03 07:28:44 -0500 )edit

1 Answer

Sort by ยป oldest newest most voted

answered 2016-09-03 13:16:25 -0500

curranw gravatar image

updated 2016-09-03 13:16:50 -0500

I have some general techniques that seem to work with haar classifiers. The same may work for you.

If I'm getting too many false positives, a great tool to use is the opencv_createsamples executable. It allows you to generate many additional positive samples by stretching, rotating and zooming in/out positives. It also throws these samples in other negative images, so your classifier learns to ignore backgrounds. Here's a good tutorial for it.

If I'm not getting enough true positives, I simply need more positives examples. These could be just brute-forcing more examples, or trying more feature-rich positive examples.

edit flag offensive delete link more


Thanks @curranw -- this might also work for face recognition so I'll give it a try.

Pi Robot gravatar image Pi Robot  ( 2016-09-05 19:32:01 -0500 )edit

Hi, Can you please help me how can I use this tutorial for cob people detection . Thanks for advance

amal gravatar image amal  ( 2018-07-12 05:34:15 -0500 )edit

Question Tools



Asked: 2016-09-02 13:08:26 -0500

Seen: 224 times

Last updated: Sep 03 '16