Superpixels and Polygons Using Simple Non-iterative Clustering

Abstract

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 Algorithm1. 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. 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.

Boundary-Aware Superpixel Segmentation Based on Minimum Spanning Tree

“…Then, Some Fig. 7 Some subjective results for compared methods, from top to bottom and left to right: MST, LSC [6], SNIC [7], SLIC [5], NC [4], TP [15], RPS [8] and FH [9]. Fig.…”

“…Boundary recall (BR), under segmentation error (UE) and achievable segmentation accuracy (ASA) [14] are used to evaluate superpixel segmentation performance. We compare seven state-of-the-arts: LSC [6], SNIC [7], SLIC [5], NC [4], TP [15], RPS [8] and FH [9]. Figure 4 shows the segmentation performance along with the increasing superpixel number for all compared methods.…”

“…The results are shown in Table 1. SNIC [7] obtains the fastest implementation due to their C++ code. Our method is comparable to the state-of-the-arts although the code is implemented in Matlab without any optimization.…”

“…The first class is Euclid-distance based similarity metric, such as Ncut [4], SLIC [5], LSC [6] and SNIC [7]. The second class is geodesic distance based similarity metric, such as RPS [8] and FH [9].…”

Texture Superpixel Clustering from Patch-based Nearest Neighbor Matching

“…2) Compared methods: NNSC performances are compared to the ones of the recent state-of-the-art methods SLIC [8], ERGC [19], ETPS [11], LSC [13], SNIC [14], SCALP [15], and TASP [17], used with parameters recommended by the authors. Performances are measured with the standard Achievable Segmentation Accuracy (ASA) [9] that evaluates the accuracy of superpixels according to a ground truth segmentation.…”