Superpixels via Multi-Label Graph-Cuts
Yuhang Zhang (ANU)
COMPUTER VISION AND ROBOTICS SERIESDATE: 2011-07-21
TIME: 16:00:00 - 17:00:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
Superpixels provide a more simplified and more meaningful representation of an image than the ordinary pixels. They are now widely used in various computer vision applications. However, superpixel segmentation by itself is not a simple process. The traditional superpixel algorithm, which recursively implements NCuts, is well known for its low efficiency. Processing a single image of moderate size usually take minutes.
In this seminar we will introduce two new graph-based methods for superpixel segmentation. The main advantage of our method is its high speed. Processing an 481 by 321 image takes only 0.5 seconds, which is more than a hundred times faster then the traditional NCuts based method, and ten times faster than the Expansion-Moves based method proposed by Veksler et al. on ECCV2010. Moreover, the efficiency of our method is independent of the number of Superpixels, and nearly linear to the size of the image.
Inspired by Veksler et al., we formulate superpixel segmentation as a multi-labelling problem as well. Whereas Veksler et al. used as many labels as the number of superpixels in their formulation; our two methods used only 4 labels and 2 labels respectively. Whereas Veksler et al. use alpha-expansion, which is an iterative method to optimize their very-many-label problem, we propose a non-iterative method to optimize our 4-label problem. Our 2-label problem can be directly solved by many existing Pseudo-Boolean optimizer.
Experiments on a benchmark dataset show that, besides efficiency our methods are comparable or even better than the previous methods in the accuracy aspects, as well.
BIO:
Yuhang Zhang is a final year PhD candidate working with Prof. Richard Hartley in the Computer Vision & Robotics Group of ANU.
