Quick Shift and Kernel Methods for Mode Seeking

日常读论文(1)

Quick shift and kernel methods for mode seeking, Vedaldi, A. and Soatto, S. European Conference on Computer Vision, 2008

Overview

In this paper we show that the computational complexity of Euclidean
medoid shift is only O(dN2) (with a small constant), which makes it
faster (not slower!) than mean shift (Sect. 3).

In summary, we show that kernels extend mode seeking
algorithms to non-Euclidean spaces in a simple, general and efficient way

In this paper we exploited kernels to extend mean shift and other mode seeking
algorithms to a non-Euclidean setting.

We also introduced quick shift, which can balance under- and over-frag-
mentation of the clusters by the choice of a real parameter. We showed that,
in practice, this algorithm is very competitive, resulting in good (and sometimes
better) segmentations compared to mean shift, at a fraction of the computation
time.

Mode-seeking algorithm comparsion

• Mean shift moves the points uphill towards the mode approximately
  following the gradien (left)
• Medoid shift approximates mean shift trajectories by
  connecting data points. (middle)
• Quick shift (Sect. 3) seeks the energy modes
  by connecting nearest neighbors at higher energy levels, trading-off mode over- and
  under-fragmentation.(right)

Application: Image segmentation

Comparing their execution time, quick shift is faster.

Skimage library-segmentation-quickshift

from skimage.segmentation import quickshift
img = "../image/test/3096.jpg"
img = io.imread(img)
segments_quick = quickshift(img, kernel_size=5, max_dist=10, ratio=0.5)