I am looking for state of the art methods to capture the human segments of images (mostly clothing images that are including only one person with natural a background, for example http://i.dailymail.co.uk/i/pix/2011/05/18/article-1388242-0C1FD2EF00000578-19_468x715.jpg ). The next stage will be the segmentation of the clothes of the person. If you know any studies touching on these topics, it would be great to know.
Hello Eren, Please go through this
Automatic human gesture recognition, and specifically the problem of gesture boundary detection in a continuous gesture sequence, has gained significant interest in recent years. Several methods have been proposed but they have their limitations.
For example, the traditional dynamic time warping approach suffers from multi-label confusion, augmenting video with audio is practical in only some circumstances, and requiring resting frames in between gestures prevents usage in natural settings. A novel method of gesture segmentation which uses motion history image templates, optical flow, and motion discontinuity to break a complete gesture into its basic components, called atomic gestures. These atomic gestures are then combined to give the most probable gesture using probabilistic models such as hidden Markov models.
This approach will take care of the multi-label gesture problem as well as properly handling gesture transitions with no resting frames in gesture sequence. In addition, better boundary detection improves classification accuracy.
Hi,
You may like to try this approach:
https://www.researchgate.net/publication/221304279_What_Where_and_How_Many_Combining_Object_Detectors_and_CRFs
The software library is here:
http://cms.brookes.ac.uk/staff/PhilipTorr/ale.htm
or here:
http://www.inf.ethz.ch/personal/ladickyl/
Our approach treats the segmentation of an object as a labelling problem, where each pixel in the image is assigned to one of a pre-set number of labels, in your case this will be person Vs background. It requires hand labelled training data, you may need to label some (around 100-to-1000 usually does the trick) representative images to get the best performance. A bounding box detector, in your case a pedestrian detector should be okay, is then incorporated to achieve maximum labelling performance. A pre-trained pedestrian detector (*The INRIA one) is available from http://www.cs.berkeley.edu/~rbg/latent/index.html.
Work specifically on people has also been carried out within similar frameworks:
Ľubor Ladický, Philip H.S. Torr, Andrew Zisserman
Human Pose Estimation using a Joint Pixel-wise and Part-wise Formulation
Conference of Computer Vision and Pattern Recognition, 2013.
Vibhav Vineet, Jonathan Warrell, Lubor Ladicky, Philip H.S. Torr
Human Instance Segmentation from Video using Detector-based Conditional Random Fields
BMVC-2011
Vibhav Vineet, Glenn Sheasby, Jonathan Warrell, Philip H.S. Torr
PoseField: An Efficient Mean-field based Method for Joint Estimation of Human Pose, Segmentation and Depth (oral)
EMMCVPR 2013, Lund, Sweden
Sorry about the self reference, but it just may work for you :)
Cheers
*edit
Conference Paper What, Where and How Many? Combining Object Detectors and CRFs
Hi Eren,
you can even use this web page to segment your images....it should distinguish clothes and skin...
cheers
eren
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing