SCALP: Superpixels with Contour Adherence using Linear Path

Back to research page


  • New image decomposition method into Superpixels with Contour Adherence using Linear Path (SCALP)
  • SCALP uses color and contour features on the linear path from pixels to superpixel barycenters
  • SCALP produces very regular superpixels but with high contour adherence, and is robust to noise
  • SCALP outperforms the state-of-the-art methods on superpixel and contour detection metrics


Iterative clustering framework
  • SCALP is based on the Simple Linear Iterative Clustering (SLIC) framework [4]
  • Initial grid partition into S×S superpixels, and clustering of pixels within 2S×2S boxes for each superpixel Sk
  • Color and spatial distances are considered to gather spatially close and homogeneous pixels so
    the clustering distance between a pixel p and a superpixel Sk is:     D(p,Sk) = dcolor(p,Sk) + dspatial(p,Sk)


Use of pixel neighborhood
  • SCALP considers the neighborhood information to produce a more accurate decomposition and be robust to noise
  • The pixel neighborhood N(p) improves the color distance:     D(p,Sk) = dcolor(N(p),Sk) + dspatial(p,Sk)

Linear path definition
  • Linear path Pk
    ={q} between the position Xp of a pixel p and the barycenter Xk of the considered superpixel Sk

Color distance on linear path
  • SCALP considers the color distance on Pk
    :     D(p,Sk) = (dcolor(N(p),Sk) + dcolor(Pk
    ) + dspatial(p,Sk)
  • The color distance on the linear path prevents irregular shapes to appear:

Contour distance on linear path
  • A contour prior map can be given to SCALP to enforce the respect of image objects.
  • The maximum of contour intensity on Pk
    is considered in a new term dcontour(Pk

  • SCALP clustering distance is defined as:   D(p,Sk) =


    dcolor(N(p),Sk) + dcolor(
    ,Sk) + dspatial(p,Sk)



Robustness to noise
  • SCALP is robust to noise, contrary to most state-of-the-art methods

A Gaussian noise has been added to the bottom-right corner of the image

Qualitative results
  • SCALP provides the most visually satisfying results with regular superpixels that adhere well to the image contours

Quantitative results
  • Comparison on contour detection (PR), respect of image objects (ASA), contour adherence (BR vs CD) and regulariy (SRC) on the BSD dataset


  • C implementation (with optional Matlab/C-Mex wrapper) of the SCALP method

    Download here

Main (to cite)

  1. R. Giraud, V.-T. Ta and N. Papadakis: Robust Superpixels using Color and Contour Features along Linear Path
    Computer Vision and Image Understanding (CVIU), 2018

  1. R. Giraud, V.-T. Ta and N. Papadakis: SCALP: Superpixels with Contour Adherence using Linear Path
    International Conference on Pattern Recognition (ICPR), 2016  
  2. R. Giraud, V.-T. Ta and N. Papadakis: Décomposition en superpixels via l’utilisation de chemin linéaire
    Groupe d'Études du Traitement du Signal et des Images (GRETSI), 2017  

  1. M. Y. Liu, O. Tuzel, S. Ramalingam and R. Chellappa: Entropy rate superpixel segmentation
    IEEE Int. Conf. on Computer Vision and Pattern Recognition (CVPR), 2011
  2. R. Achanta, A. Shaji, K. Smith, A. Lucchi, et al.: SLIC superpixels compared to state-of-the-art superpixel methods
    IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2012
  3. M. Van den Bergh, X. Boix, G. Roig, B. de Capitani and L. Van Gool: SEEDS: Superpixels extracted via energy-driven sampling
    European Conference on Computer Vision (ECCV), 2012
  4. P. Buyssens, I. Gardin, S. Ruan and A. Elmoataz: Eikonal-based region growing for efficient clustering
    Image and Vision Computing, 2014
  5. J. Yao, M. Boben, S. Fidler and R. Urtasun: Real-time coarse-to-fine topologically preserving segmentation
    IEEE Int. Conf. on Computer Vision and Pattern Recognition (CVPR), 2015
  6. J. Chen, Z. Li and B. Huang: Linear spectral clustering superpixel
    IEEE Trans. on Image Processing, 2017

Back to top