Segmentation semantique profonde par regression sur cartes de distances signees

Published in Conference Reconnaissance des Formes, Images, Apprentissage et Perception, RFIAP, 2018

Authors: N. Audebert, A. Boulch, B. Le Saux, S. Lefevre


Page at HAL

Download paper here


Understanding of visual scenes relies more and more on dense pixel-wise classification obtained via deep fully convolutional neural networks. However, due to the nature of the networks, predictions often suffer from blurry boundaries and ill-segmented shapes, fueling the need for post-processing. This work formulates the standard semantic segmentation problem in terms of distance regression. We show that it is possible to train a multi-task fully convolutional neural network that builds more regular segmentations than those produced by existing methods based on direct dense classification.


  TITLE = { { Segmentation s{\'e}mantique profonde par r{\'e}gression sur cartes de distances sign{\'e}es } },
  AUTHOR = {Audebert, Nicolas and Boulch, Alexandre and Le Saux, Bertrand and Lef{\`e}vre, S{\'e}bastien},
  URL = {},
  BOOKTITLE = { { Reconnaissance des Formes, Image, Apprentissage et Perception (RFIAP)} },
  ADDRESS = {Marne-la-Vall{\'e}e, France},
  YEAR = {2018},
  MONTH = Jun,
  KEYWORDS = {Semantic segmentation ; deep learning ; distance transform ; Segmentation s{\'e}mantique ; cartes de distance ;  apprentissage profond},
  PDF = {},
  HAL_ID = {hal-01809991},
  HAL_VERSION = {v1},