Advanced

Instance-level Semantic Segmentation by Deep Normalized Cuts

Davidson, Alexander LU (2017) In Master's Theses in Mathematical Sciences FMA820 20162
Mathematics (Faculty of Engineering)
Abstract
Instance-level semantic segmentation refers to the task of assigning each pixel in an image an object class and an instance identity label. Due to its complexity, this problem has received little attention by the research community in the past. However, owing to the recent successes of deep learning technology, solving it suddenly seems possible. Starting from an initial pixel-level semantic segmentation, we combine ideas from spectral clustering and deep learning to investigate the suitability of using normalized cuts for additional instance-level segmentation. Specifically, the similarity matrix on which cuts are based is learned from examples using a convolutional network. We use the Cityscapes dataset for urban street scene... (More)
Instance-level semantic segmentation refers to the task of assigning each pixel in an image an object class and an instance identity label. Due to its complexity, this problem has received little attention by the research community in the past. However, owing to the recent successes of deep learning technology, solving it suddenly seems possible. Starting from an initial pixel-level semantic segmentation, we combine ideas from spectral clustering and deep learning to investigate the suitability of using normalized cuts for additional instance-level segmentation. Specifically, the similarity matrix on which cuts are based is learned from examples using a convolutional network. We use the Cityscapes dataset for urban street scene understanding to train our deep normalized cuts model to segment individual cars. When using the ground truth number of instances to help determine the correct number of cuts, our method attains almost twice the performance of our connected components baseline. (Less)
Please use this url to cite or link to this publication:
author
Davidson, Alexander LU
supervisor
organization
course
FMA820 20162
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Deep learning, Machine learning, Convolutional neural networks, Normalized cuts, Spectral clustering, Instance-level semantic segmentation
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3329-2017
ISSN
1404-6342
other publication id
2017:E55
language
English
id
8925933
date added to LUP
2017-10-25 15:46:00
date last changed
2017-10-25 15:46:00
@misc{8925933,
  abstract     = {Instance-level semantic segmentation refers to the task of assigning each pixel in an image an object class and an instance identity label. Due to its complexity, this problem has received little attention by the research community in the past. However, owing to the recent successes of deep learning technology, solving it suddenly seems possible. Starting from an initial pixel-level semantic segmentation, we combine ideas from spectral clustering and deep learning to investigate the suitability of using normalized cuts for additional instance-level segmentation. Specifically, the similarity matrix on which cuts are based is learned from examples using a convolutional network. We use the Cityscapes dataset for urban street scene understanding to train our deep normalized cuts model to segment individual cars. When using the ground truth number of instances to help determine the correct number of cuts, our method attains almost twice the performance of our connected components baseline.},
  author       = {Davidson, Alexander},
  issn         = {1404-6342},
  keyword      = {Deep learning,Machine learning,Convolutional neural networks,Normalized cuts,Spectral clustering,Instance-level semantic segmentation},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences},
  title        = {Instance-level Semantic Segmentation by Deep Normalized Cuts},
  year         = {2017},
}