|
What are some of the top performers in image segmentation? If you can make some good recommendations for reading material, that's a big plus. I've surveyed a few approaches/algorithms based on edge-detection techniques, as well as a few based on clustering and local histograms, but I haven't found much in the way of summary information comparing & contrasting the various algorithms out there. Thanks in advance. |
|
Maybe one addition: If you are looking for low level segmentation, the following approaches are quite popular. They don't provide state of the art performance as gPb-based methods do but are used quite a lot: - Mean shift - Quick shift - Normalized Cuts - Felsenzwalb's fast segmentation - Watershed - CRF-Based I don't know of a comparison of these methods that really convinced me. There is a paper on classification using superpixels that compares the different approaches to obtain superpixels: http://www.prip.tuwien.ac.at/people/hanbury/files/hanbury_superpixels_segmentation.pdf This is a pretty specific setup, though. There is also work on "SLIC", which is just K-Means on the (5d) pixel values. They claim to be the best but I'm pretty sure their approach breaks down if the segments are more than just a few pixels wide.
This answer is marked "community wiki".
|
|
Hello, I might be a bit of help here. In the recent summer School on Machine Learning in Singapore, Max Welling and Stephen Gould gave exellent lectures on object recognition via graph cutting using Markov Random Fields. The tutorials are not online, but they did prepare excellent self contained slides. These are the ones of Max Welling. These are the ones of Stephen Gould. They are in three parts, One, Two, Three |
|
What do you mean by image segmentation? Most current work focuses on object class segmentation. If you want I can recommend lot's of stuff about that ;) A good starting point there is the pascal voc challenge. If you want more low level segmentation (such as is done for the berkeley segmentation challenge) you should look at the work of Jitendra Malik's group. The "global probablility of boundary" detector is a standard approach by now and there are some papers continuing this work. If you can specify what problem you are interested in, I can give you more infos. The problem I'm actually working on is only tangential to image segmentation. I'm looking for a relatively decent way to identify a potential object of interest to be cut out of the scene. It doesn't necessarily have to be separated from the surrounding image, it could just be a rectangular cut-out roughly centered on whatever it found in the scene.
(Dec 07 '11 at 11:41)
Brian Vandenberg
What do you mean by "object of interest"? Is this a specific category or something along the lines of "saliency"? Detecting bounding boxes of object belonging to specific semantic categories is a well-established task and there is much research in this direction. But I'm not sure if that is what you are looking for. There is work by Itti on finding salient object ins scenes, not sure if that matches your idea of "object of interest". There is also the paper what is an object. Don't know if that is what you are looking for.
(Dec 08 '11 at 03:49)
Andreas Mueller
For example, if I take a picture with a few items on a table, extracting cut-outs of the individual objects. I'll look into the links you posted this weekend when I get a little bit of free time.
(Dec 09 '11 at 12:12)
Brian Vandenberg
1
Ok this sounds not class specific. Then you should definitely look at this page: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/index2.html Maybe start with this one: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/papers/amfm_pami2010.pdf
(Dec 09 '11 at 13:04)
Andreas Mueller
|
From reading papers in different areas it seems like some form of Markov random fields, edge detections, and superpixels are usually used for this. Keywords are low-level vision, markov random fields image segmentation, markov random fields low-level-vision, etc. I can't really be more specific because I haven't read the papers that come out when I search for this.
Have any of the algorithms you've read about stood out to you? Can you think of any in particular you found intriguing or impressive?
Using some of the key-words Alex mentioned, I found a paper that looks rather promising -- more in the vein of unsupervised approaches (which is what I was hoping to find): "Multiscale Image Segmentation by Integrated Edge and Region Detection", by Mark Tabb & Narendra Ahuja.