Geodesic Propagation for Semantic Labeling

Xiaowu Chen  , Qing Li  , Yafei Song, Jin Xin1, and Qinping Zhao

1Beijing Electronic Science and Technology Institute, Beijing 100070, China

Abstract

    In this paper we propose a novel semantic label transfer method using supervised geodesic propagation (SGP). We use supervised learning to guide the seed selection and the label propagation. Given an input image, we first retrieve its similar image set from annotated databases. A Joint Boost model is learned on the similar image set of the input image. Then the recognition proposal map of the input image is inferred by this learned model. The initial distance map is de- fined by the proposal map: the higher probability, the smaller distance. In each iteration step of the geodesic propagation, the seed is selected as the one with the smallest distance from the undetermined superpixels. We learn a classifier as an indicator to indicate whether to propagate labels between two neighboring superpixels. The training samples of the indicator are annotated neighboring pairs from the similar image set. The geodesic distances of its neighbors are updated according to the combination of the texture and boundary features and the indication value. Experiments on three datasets show that our method outperforms the traditional learning based methods and the previous label transfer method for the semantic segmentation work.

Gallery

geodesic-1

Our objective is to get the semantic labeling of the input image. This figure shows our results on CamVid dataset and MSRC dataset. Best viewed in color.

geodesic-2

The pipeline of our framework. Given the input image, we infer its proposal map using the boosted model learned on the annotated dataset. The initial seeds for geodesic propagation are selected based on the proposal map. The initial geodesic distance map is defined as: the higher probability, the smaller distance. With the computed weights based on features extracted from the image, the geodesic distance is updated during propagation. Finally the labels of seeds are propagated to the rest image pixels.

geodesic-3

The workflow of SGP algorithm. Given the input image, we get its similar image set from the annotated dataset using Gist retrieval (SectionV-A). The boosted model is trained on this similar image set. Besides, the propagation indicator implying the contextual information is trained on the similar image set as well (Section V-C).

geodesic-4

The pipeline of our hybrid geodesic propagation for video. Given the key frames as well as their label annotations, we propagate the semantic labels throughout the whole video. In both spatial and temporal propagation,geodesic distance is exploited to propagate accurate label.

geodesic-5

Our GP results on the four datasets.

geodesic-6

Our SGP results on the four datasets.

geodesic-7

The result of video propagation. These frames are from CamVid sequence Seq06R0. Semantic labels are overlaid on the image.

Paper

pdf
, , , , , :Geodesic Propagation for Semantic Labeling. IEEE Trans. Image Processing 23(11): 4812-4825 ().
pdf
, , , , :Supervised Geodesic Propagation for Semantic Label Transfer. ECCV (3) : 553-565.

BibTeX

@article{DBLP:journals/tip/LiCSZJZ14,
  author    = {Qing Li and
               Xiaowu Chen and
               Yafei Song and
               Yu Zhang and
               Xin Jin and
               Qinping Zhao},
  title     = {Geodesic Propagation for Semantic Labeling},
  journal   = {{IEEE} Trans. Image Processing},
  volume    = {23},
  number    = {11},
  pages     = {4812--4825},
  year      = {2014},
  url       = {https://doi.org/10.1109/TIP.2014.2358193},
  doi       = {10.1109/TIP.2014.2358193},
  timestamp = {Fri, 26 May 2017 22:51:39 +0200},
  biburl    = {http://dblp.uni-trier.de/rec/bib/journals/tip/LiCSZJZ14},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}
@inproceedings{DBLP:conf/eccv/ChenLSJZ12,
  author    = {Xiaowu Chen and
               Qing Li and
               Yafei Song and
               Xin Jin and
               Qinping Zhao},
  title     = {Supervised Geodesic Propagation for Semantic Label Transfer},
  booktitle = {Computer Vision - {ECCV} 2012 - 12th European Conference on Computer
               Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part {III}},
  pages     = {553--565},
  year      = {2012},
  crossref  = {DBLP:conf/eccv/2012-3},
  url       = {https://doi.org/10.1007/978-3-642-33712-3_40},
  doi       = {10.1007/978-3-642-33712-3_40},
  timestamp = {Fri, 19 May 2017 01:25:56 +0200},
  biburl    = {http://dblp.uni-trier.de/rec/bib/conf/eccv/ChenLSJZ12},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}

发表评论

电子邮件地址不会被公开。 必填项已用*标注