(Locally Debiased) Region Contrast Saliency
This package provides the implementation of the locally debiased region contrast saliency algorithm. Furthermore, it also provides the original region contrast algorithm and allows to set different optional center bias integration schemes (min, max, linear, product).
Locally debiased region contrast saliency has the advantage that it is not center biased, which makes it a perfect candidate as a salient object detection algorithm for, e.g., surveillance footage or robots (i.e., image data that was not collected/generated by a photographer). However, you can also bias the algorithm again, in which case the algorithm also provides state-of-the-art performance on Achanta's salient object detection dataset.
If you use any of this work in scientific research or as part of a larger software system, you are kindly requested to cite the use in any related publications or technical documentation. The work is based upon:
B. Schauerte, R. Stiefelhagen, "How the Distribution of Salient Objects in Images Influences Salient Object Detection". In Proceedings of the 20th International Conference on Image Processing (ICIP), 2013.
We applied the region contrast algorithm to learn better attribute (color term) models from images collected from the Web by focusing on the salient object, see
B. Schauerte, R. Stiefelhagen, "Learning Robust Color Name Models from Web Images". In Proceedings of the 21st International Conference on Pattern Recognition (ICPR), 2012
Furthermore, the debiased region contrast has been applied in:
Schauerte, R. Stiefelhagen, "Look at this! Learning to Guide Visual Saliency in Human-Robot Interaction". In Proceedings of the International Conference on Intelligent Robots and Systems (IROS), 2014.
The code can be compiled using the build.m script. However, tt might be necessary to change some of the pathes in the script, depending on your system configuration.
The code has been successfully compiled under Mac OS X and Linux.
Threee demonstrations/visualizations are included
- test_region_saliency_mex.m calculates the visual saliency for several included algorithms. This is the best starting point, if you want to play with the code.
- visualize_region_contrast_distance_bias.m demonstrates the bias in the original region contrast algorithm.
- visualize_theoretical_distance_bias.m gives an overview of how different metrics lead to different biases.
4. AUTHOR INFORMATION
4.2 CODE HOSTING
Development version, feel free to commit: LDRC@GitHub
Part of this work was/is supported by the German Research Foundation (DFG) within the Collaborative Research Program SFB 588 "Humanoid Robots" and the Quaero Programme, funded by OSEO, French State agency for innovation.