Stochastic Image Reconstruction from
Local Histograms of Gradient Orientation

Presentation   -   Papers   -   Source codes   -   Credits

Presentation

  In this work, we tackle the problem of image reconstruction from local descriptors. More precisely, we propose two stochastic models (MS-Poisson and MaxEnt) which allow to explore the set of images having similar descriptors. These models are designed to reconstruct from simplified SIFT descriptors, i.e., histogram of oriented gradients (HOG) computed in the subcells associated to SIFT keypoints. But the reconstruction method can also applied to the true SIFT descriptors. These models are able to recover global shapes and many geometric details of images. They compare well to state of the art results, without requiring the use of any external database.

Papers

  Our reconstruction algorithm is described in the papers

"Stochastic Image Models from SIFT-like descriptors" (Agnès Desolneux, Arthur Leclaire),
submitted to the SIAM Journal on Imaging Sciences, 2018. Preprint.

"Stochastic Image Reconstruction from Local Histograms of Gradient Orientation"
(Agnès Desolneux, Arthur Leclaire), in the Proceedings of the International Conference on Scale Space and Variational Methods in Computer Vision, 2017.
Revised Preprint.

  The experiments shown in this paper can be reproduced using the codes available below.

Source codes (Matlab)

  The archive release.zip contains source codes for image reconstruction.

  This archive contains the main functions

• reconstruction_mspoisson.m : reconstruction from simplified SIFT descriptors with MS-Poisson,
• reconstruction_sift.m : reconstruction from SIFT descriptors with MS-Poisson,
• reconstruction_maxent.m : reconstruction from local HOG with MaxEnt,
• reconstruction_hog.m : reconstruction from dense HOG with Poisson,

  and also

• a script file tutorial.m,
• a script file expe_article_journal.m which allows to reproduce the experiments of our papers
• miscellaneous functions needed by the reconstruction methods
• a folder containing input images and corresponding SIFT keypoints used in our experiments.

  You can also use these reconstrution functions for other images but then you should compute the SIFT keypoints and descriptors. For that you may use David Lowe's Matlab implementation of the SIFT method available here.

  If you find some bugs or mistakes in those codes, you can report them by email to Arthur Leclaire.

Credits

  We would like to thank the researchers who provided the images used in our experiments.

  The images vondrick_* were taken from the article

Hoggles: Visualizing object detection features. (C. Vondrick, A. Khosla, T. Malisiewicz, and A. Torralba), Proceedings of the IEEE ICCV, 2013.

  The images perez_* were taken from the article

Reconstructing an image from its local descriptors. (P. Weinzaepfel, H. Jégou, and P. Pérez), Proceedings of the IEEE CVPR, 2011.