|
Face detection
Anonymous from United States [4 posts] |
10 year
|
|
I know a lot of advances have been made in the area of software face detection, but it seems to me it's still difficult for a computer to recognize a face turned sideways. I was wondering if anyone has had some experience using RoboRealm to detect a face in profile?
|
|
|
|
JD Martin from United States [1 posts] |
10 year
|
You might find some answers on forums.usfirst.org. I've seen support for RoboRealm applications there for Axis cameras.
It's an interesting problem. I myself am not sold on the use of face recognition software for anything other than a starting point, i.e. to sift through thousands of images to reduce it to a manageable few for humane inspection.
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
There have been some studies done on profile detection ... but the techniques used for face detection require a large number of images to be annotated and then trained against. Even in ours, we had annotated about 100 images which is much less than other systems but still very laborious to do. If you happen to have about 100 profile images that we can use to train the system on, that's certainly possible.
Otherwise, that's probably a good summer intern project! :-)
STeven.
|
|
|
|
Anonymous from United States [4 posts] |
10 year
|
|
That's an interesting possibility, Steven. I don't need face recognition - I need an application that can look at a still image or video file and be able to detect that a human face, even if turned sideways, is present in the picture. Would you need a variety of pictures of faces turned in different positions, etc., in order to train the program?
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
Yes, typically for these types of applications you need training samples (and some testing too) such that the learning algorithm can use the entire set of training examples to determine unique features about what needs to be detected. This is an automated process but training requires that the application *knows* what within any image it needs to identify. So, you need many images with an annotation (i.e. draw a red square around each face/profile) in approximately the same way so that prior to training the system can automatically extract out the 'correct' answer and train the system accordingly. You will need two images, one with the annotation and one without as drawing in the image changes it which makes it unusable for training.
See the two attached images as an example. First is the raw unchanged image, the second is with the annotation.
Its important to use different images in different conditions. When we trained our face detection we couldn't get many images with people with glasses and thus the system doesn't detect faces with glasses very well!
There are many issues that you have to be aware of depending on which images you used. One can use google images or the like to find faces but things like focus all the way through as apposed to just on the face are important (i.e. you need to include both). Lighting is also important .. i.e. don't just use images of models that are typically illuminated more than the background, etc.
If you are serious about undertaking such an adventure let me know ... we have an annotation tool that we used for the face detection that is overdue for some upgrades that make annotation a bit quicker ... but you still need to spend the time collecting images (flicker, google images, facebook, etc. are all good sources). Note that we don't publish the images we use for training nor include any original data from those images so using images that you cannot display/publish/etc. should not be an issue.
STeven.
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
Images attached.
 
|
|
|
|
Anonymous from United States [4 posts] |
10 year
|
|
I might be interested depending on how much spare time I have; I will be contacting you. Please check the email from your contact form. Thanks.
|
|
This Thread
