TY - JOUR
T1 - Superpixel Segmentation with Fully Convolutional Networks
AU - Yang, Fengting
AU - Sun, Qian
AU - Jin, Hailin
AU - Zhou, Zihan
N1 - Funding Information:
Thispaperhaspresentedasimplefullyconvolutional networkforsuperpixelsegmentation. Experimentson benchmark datasets show that the proposed model is computationally efficient, and can consistently achieve the state-of-the-artperformancewithgoodgeneralizability. Further, we havedemonstrated that higherdisparity estimation accuracy can be obtained by using superpixels to preserve object boundaries and fine details in a popular stereo match-ingnetwork. Inthefuture, weplantoapplythepro-posedsuperpixel-baseddownsampling/upsampling scheme tootherdensepredictiontasks,suchasobjectsegmentation andopticalflowestimation, andexploredifferent waysto use superpixels in these applications. Acknowledgement. This work is supported in part by NSF award #1815491 and a gift from Adobe.
Publisher Copyright:
© 2020 IEEE.
PY - 2020
Y1 - 2020
N2 - In computer vision, superpixels have been widely used as an effective way to reduce the number of image primitives for subsequent processing. But only a few attempts have been made to incorporate them into deep neural networks. One main reason is that the standard convolution operation is defined on regular grids and becomes inefficient when applied to superpixels. Inspired by an initialization strategy commonly adopted by traditional superpixel algorithms, we present a novel method that employs a simple fully convolutional network to predict superpixels on a regular image grid. Experimental results on benchmark datasets show that our method achieves state-of-the-art superpixel segmentation performance while running at about 50fps. Based on the predicted superpixels, we further develop a downsampling/upsampling scheme for deep networks with the goal of generating high-resolution outputs for dense prediction tasks. Specifically, we modify a popular network architecture for stereo matching to simultaneously predict superpixels and disparities. We show that improved disparity estimation accuracy can be obtained on public datasets.
AB - In computer vision, superpixels have been widely used as an effective way to reduce the number of image primitives for subsequent processing. But only a few attempts have been made to incorporate them into deep neural networks. One main reason is that the standard convolution operation is defined on regular grids and becomes inefficient when applied to superpixels. Inspired by an initialization strategy commonly adopted by traditional superpixel algorithms, we present a novel method that employs a simple fully convolutional network to predict superpixels on a regular image grid. Experimental results on benchmark datasets show that our method achieves state-of-the-art superpixel segmentation performance while running at about 50fps. Based on the predicted superpixels, we further develop a downsampling/upsampling scheme for deep networks with the goal of generating high-resolution outputs for dense prediction tasks. Specifically, we modify a popular network architecture for stereo matching to simultaneously predict superpixels and disparities. We show that improved disparity estimation accuracy can be obtained on public datasets.
UR - http://www.scopus.com/inward/record.url?scp=85094819681&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85094819681&partnerID=8YFLogxK
U2 - 10.1109/CVPR42600.2020.01398
DO - 10.1109/CVPR42600.2020.01398
M3 - Conference article
AN - SCOPUS:85094819681
SN - 1063-6919
SP - 13961
EP - 13970
JO - Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
JF - Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
M1 - 9156320
T2 - 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020
Y2 - 14 June 2020 through 19 June 2020
ER -