BingObjCVPR14 source code readme

Download the VOC 2007 data (2 zip files for training and testing) and put them in ./VOC2007/

This program needs OpenCV readable data. An OpenCV readable of VOC 2007 annotation could be download from:

After unzip all the data package, put them in the same folder. Some files or folders from these data packages might have the same name, feel free to merge folders with the same name and overwrite files with the same name, in any order. If you put the data in the default VOC2007 folder, open the project with Visual studio 2012, and have latest opencv installed properly, you should be able to run the code by press Ctrl+F5 key together.


If you use any part of our paper in your research, please cite our paper:

  • BING: Binarized Normed Gradients for Objectness Estimation at 300fps. Ming-Ming Cheng, Ziming Zhang, Wen-Yan Lin, Philip Torr, IEEE CVPR, 2014. [Project page][pdf][bib][C++]

You might be also interested in:

  • Global Contrast based Salient Region Detection. Ming-Ming Cheng, Guo-Xin Zhang, Niloy J. Mitra, Xiaolei Huang, Shi-Min Hu. IEEE CVPR, 2011, p. 409-416. [Bib] [Pdf 15M] [Pdf 中文版] [C++] [FAQs]  (#2 most cited paper in CVPR 2011)
  • Salient Object Detection and Segmentation. Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, Shi-Min Hu. Submitted to IEEE TPAMI (TPAMI-2011-10-0753), 2011. [Pdf] [Poster] [Bib]

This code is extremely fast, thus many small issues, which typically can be ignored, might influence counting the time.

  1. Hard disk speed. The time of reading images from hard disk might not be ignorable. Uncomment the //#define PRE_LOAD_IMAGE in objectness.cpp to enable preloading all the image to memory, thus ignore time counting for image read. To uncomment this line, it requires sufficient RAM space.
  2. Running in release mode is VERY important to get extreme performance when never STL is used, e.g. std::vector, std::map, etc.
  3. The source code was compiled in Visual Studio 2012, 64 bit Windows systems, and OpenCV2.4.8.  If you have made a version running on other platforms (Software at other platforms, e.g. Mac, Linux, vs2010, makefile projects) and want to share it with others, please send me an email containing the URL and I will add a link in the project page:
  4. There are some random sampling process for getting training samples. Each time running the code, the results might be slightly different. Statistically the recall given 1000 proposals typically in the range 95.8% ~ 96.4%.

If you have any problem with running the source code, please read the FAQs in the project page before sending me emails.