Gao et al. propose saliency-guided supervoxels (SSV), a supervoxel algorithm improving over Voxel cloud connectivity segmentation (VCCS) by adaptively choosing the supervoxel resolution based on saliency. VCCS is a supervoxel algorithm operating directly within the point cloud obtained form a RGBD image, see [1] for details. Given a saliency map, in their case computed by VOCUS2 [6]. The pixels are clustered into $K$ clusters using $k$-means. Then, the average saliency per cluster can easily be calculated. The size of the supervoxels is adapted to the individual clusters. In the case of VCCS this can be done by adapting the seed resolution (see the reading notes above). The proposed approach works with a minimum and maximum seed resolution $R_{\text{min}}$ and $_{\text{max}}$ respectively. A step size $d$ is determined as
The resolution $r_k$ for the $k$-the cluster is then given by
$r_k = 10^{\log R_{\text{max}} – (k – 1)d}$
As a result, higher saliency results in higher density of supervoxels. VCCS is run for each cluster independently and the different supervoxel segmentations are then combined.
In experiments, they report improved performance regarding Boundary Recall and Undersegmentation Error on the NYUV2 dataset [18] and the SUNRGBD dataset [19]. The parameters used for evaluation are reported. A qualitative example is shown in Figure 1.
Figure 1: Qualitative results on the image (left) for VCCS (middle) and SSV (right).
[1] Jeremie Papon, Alexey Abramov, Markus Schoeler, Florentin Wörgötter. Voxel Cloud Connectivity Segmentation - Supervoxels for Point Clouds. CVPR, 2013.
[18] N. Silberman, D. Hoiem, P. Kohli, and R. Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, 2012.
[19] S. Song, S. P. Lichtenberg, and J. Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In CVPR, June 2015.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.
Gao et al. propose saliency-guided supervoxels (SSV), a supervoxel algorithm improving over Voxel cloud connectivity segmentation (VCCS) by adaptively choosing the supervoxel resolution based on saliency. VCCS is a supervoxel algorithm operating directly within the point cloud obtained form a RGBD image, see [1] for details. Given a saliency map, in their case computed by VOCUS2 [6]. The pixels are clustered into $K$ clusters using $k$-means. Then, the average saliency per cluster can easily be calculated. The size of the supervoxels is adapted to the individual clusters. In the case of VCCS this can be done by adapting the seed resolution (see the reading notes above). The proposed approach works with a minimum and maximum seed resolution $R_{\text{min}}$ and $_{\text{max}}$ respectively. A step size $d$ is determined as
$d = \frac{ \log R_{\text{min}} - \log R_{\text{max}}}{K – 1}$
The resolution $r_k$ for the $k$-the cluster is then given by
$r_k = 10^{\log R_{\text{max}} – (k – 1)d}$
As a result, higher saliency results in higher density of supervoxels. VCCS is run for each cluster independently and the different supervoxel segmentations are then combined.
In experiments, they report improved performance regarding Boundary Recall and Undersegmentation Error on the NYUV2 dataset [18] and the SUNRGBD dataset [19]. The parameters used for evaluation are reported. A qualitative example is shown in Figure 1.
Figure 1: Qualitative results on the image (left) for VCCS (middle) and SSV (right).