SNIC makes two important modifications to SLIC :
1. Centroids are evolved using online averaging.
2. Label assignment is achieved using a priority queue, which returns
the element with the shortest distance D to a centroid.
Polygon Partitioning Algorithm
1. Segment image. Trace superpixel
boundaries using a standard algorithm.
2. Assign initial vertices to be pixels that
touch at least three different segments, at
least two segments and the image borders,
or are image corners.
3. New vertices are added using the
Douglas-Peucker curve simplification
algorithm.
4. Merge vertices that are too close and join
remaining vertices to obtain polygons.
1. Pick seeds on a regular square grid.
2. Initialize priority queue Q with immediate neighbors of seeds.
While Q is not empty:
3. Pop Q, and label the pixel P.
4. Update corresponding centroid.
5. For all unlabeled neighbors of P, compute D and push on Q.
|Q| = 16
+16
2. For each seed compute
distance D to unlabeled
neighbors and push on Q.
1
|Q| = 15
3. Pop the top-most element
on the queue and label the
corresponding pixel.
4. Compute distance D to the
nearest neighbors of this newly
labeled pixel and push on Q.
Continue until Q is empty.
|Q| = 18
+3
1. Initial seeds with a unique
label. Q is empty at this time.
|Q| =0
Unlabeled pixel
Labeled pixel
Simple Non-Iterative Clustering (SNIC) is an improved version of the
Simple Linear Iterative Clustering* (SLIC) algorithm. SNIC is non-
iterative, enforces connectivity from the start, requires less memory, is
faster, and yet is a simpler algorithm. On segmentation benchmarks
SNIC performs better than the state-of-the-art, including SLIC.
s ⇥ s
s ⇥ s
2s ⇥ 2s
Local k-means (SLIC)
Shortcomings of SLIC:
1. Several iterations
2. Repeat computations in overlapping local regions
3. Pixel connectivity enforced as a post-processing step
SLIC review
D =
kx
j
x
j
k
2
2
s
+
kc
j
c
k
k
2
2
m
c =[l, a, b]
T
x =[x, y]
T
s =
r
N
K
m = 10
SLIC performs k-means clustering on the image plane with centroids chosen
on a square grid in the image plane and distance D to be a weighted sum of
the normalized spatial and color distances.
Superpixels and Polygons using
Simple Non-Iterative Clustering
RADHAKRISHNA ACHANTA & SABINE SÜSSTRUNK
Simple Non-Iterative Clustering (SNIC) Algorithm
SNIC superpixels SNIC polygons
IVRL (IC), EPFL
Global k-means
* SLIC Superpixels Compared to the State-of-the-art Superpixel Methods.
R. Achanta, S. Shaji, K. Smith, A. Lucchi, P. Fua. S. Süsstrunk (TPAMI 2012).
200 400 600 800 1000 1200 1400 1600 1800 2000
Number of superpixels
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
Segmentation error (CUSE)
NCUTS
MST
TURBO
SLIC
SEEDS
ERS
LSC
SNIC
CONPOLY
SNICPOLY
200 400 600 800 1000 1200 1400 1600 1800 2000
Number of superpixels
0.28
0.29
0.3
0.31
0.32
0.33
0.34
0.35
0.36
0.37
0.38
F-measure
NCUTS
MST
TURBO
SLIC
SEEDS
ERS
LSC
SNIC
CONPOLY
SNICPOLY
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9
Boundary recall
0.6
0.65
0.7
0.75
0.8
0.85
0.9
Boundary precision
NCUTS
MST
TURBO
SLIC
SEEDS
ERS
LSC
SNIC
CONPOLY
SNICPOLY