the-grid/gmr-saliency

GMR Saliency Map

GMR Saliency Map

C++ implementation for paper "Saliency Detection via Graph-Based Manifold Ranking" by Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan and Ming-Hsuan Yang. To appear in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2013), Portland, June, 2013.

This implementation was written originally by Chuan Yang [email protected] (3-clause BSD license) and uses an also open source SLIC implementation written by Vilson Vieira/The Grid [email protected].

It was ported and tested on MacOS X 10.9.5 and Ubuntu Linux 14.04 using GNU C++ Compiler and OpenCV 2.4.9..

Depencencies

• OpenCV

Using

In a NodeJS environment:

npm install
grunt
./build/Release foo.png

With common C++ compiler:

make
./gmr-saliency foo.png

You can also import gmr-saliency into your NoFlo project and use the GetSaliency component to extract saliency maps from JPG/PNG image files or canvas while in browser.

gmr-saliency's Issues

Safe check all measurements extracted

Investigate node-pre-gyp

Using node-pre-gyp to precompile binaries for multiple platforms seems interesting, avoiding waiting times on CI and making building on multiplatforms much easier.

References:

• http://cylonjs.com/blog/2014/11/19/creating-multiplatform-precompiled-binaries-for-node-modules/
• lovell/sharp#186

Clean up result

Now we have the following as result:

saliency:
  bounding_rect: <array of arrays as points> box around biggest salient region
  polygon: <array of arrays as points> approximate polygon around biggest salient region
  center: <array as point> center of biggest salient region
  radius: <number> radius of the approx enclosing circle around biggest salient region

We should have the following, preserving the original fields for a while:

saliency:
  bbox: <array of objects> box around all regions
  confidence: <number> inverse/normalized entropy of saliency map histogram
  regions: <array of objects> [ 
    bbox: <array of objects> box around salient region
    polygon: <array of objects> polygon around salient region
    center: <object> center point of salient region
    radius: <number> radius of the approx enclosing circle around salient region
  ]

In this way, we should remove saliency.bounding_rect, saliency.polygon, saliency.center and saliency.radius because a saliency is composed by many regions. The only information regarding all those regions is the saliency.bbox that encloses them. If one wants to draw or use polygon of the saliency, it should iterate over all saliency.regions's bboxes.

We also should have a standard object to bboxes ({x: ?, y: ?, width: ?, height: ?}) and points ({x: ?, y: ?}).