Object Description Based on Spatial Relations between Level-Sets

MickaŽl Garnier, Thomas Hurtut, Laurent Wendling,

in Proc. IEEE Digital Imaging Computing: Techniques and Applications (DICTA), 2012

Left, an image is first decomposed along its N level-sets, then described by all the Force-Histograms between the pairs of level-sets. Middle, a Force-Histogram measures the spatial relation between two level-sets along a direction theta as their relative gravitational force along theta. Right, using a similarity distance on the proposed feature description, we classify butterflies along their species. Butterflies are indeed highly discriminated by the spatial organisation of their inner structures (spots, inner contours, etc.).


Object recognition methods usually rely on either structural or statistical description. These methods aim at describing different types of information such as the outer contour, the inner structure or texture effects. Comparing two objects then comes down to averaging different data representations which may be a tricky issue. In this paper, we introduce an object descriptor based on the spatial relations that structures object content. This descriptor integrates in a single homogeneous representation both shape information and relative spatial information about the object under consideration. We use this description in the context of image retrieval and show results on a butterfly image database compared with both GFD and dense SIFT descriptors. These results show that our method is more efficient to distinguish the objects where the spatial organization is a discriminative feature.


  author       = "Garnier, Mickael and Hurtut, Thomas and Wendling, Laurent",
  title        = "Object Description Based on Spatial Relations between Level-Sets",
  booktitle    = "IEEE Digital Imaging Computing: Techniques and Applications (DICTA)",
  year         = "2012",
  pages        = ""


Paper pdf
Database zip - www


This work was supported by the ANR SPIRIT (ANR-11-JCJC-008-01).