Software
Algorithms for Kernel-Based Efficient Subwindow Search
This is an implementation of subwindow search methods for object detection and localization that employs the SMAWK algorithm to support distance kernels for comparing feature histograms (rather than requiring the features be weighted):
- ESS from Lampert and Blaschko (CVPR 2008) modified to handle distance kernels.
- Alternating Row-Column Search (ARCS) (CVPR 2010) (formerly named A-ESS).
Additionally, a couple of other ESS-like algorithms are provided:
- Improved-ESS from An et al (CVPR 2009) modified to handle distance kernels.
- Bentley's algorithm, also modified to handle distance kernels.
Distance kernels provided are Chi-Squared and Weighted-Chi-Squared, although other histogram distance measures can easily be substituted in. The code will compile for Visual Studio 2005/2008 (a project is included in the .zip), and the source code should compile under Linux and probably under any platform supported by GCC.
Download source [KernelFastESS_v1.0.zip]
Algorithms for Efficient Subwindow Search
This is an implementation of four subwindow search methods for object detection and localization: Bentley's algorithm, ESS, Improved-ESS and Alternating-ESS. It will compile for Visual Studio 2005 (a project is included in the .zip), and the source code can compile under Linux and probably under any platform supported by GCC.
Download source [FastESS_v1.0.zip]
Ground-truth data for PETS 2007:
This is the ground-truth data manually produced by our research group for the evaluation of tracking algorithms in challenging environment. The ground-truth data is provided freely for research-only purposes. If you use this data for published work, please kindly acknowledge/cite the PETS 2007 dataset and this technical report.
Download ground-truth
Part 1: [ PETS2007-S8-Ground-Truth-Part1.tar.gz (30MB) ]
Part 2: [ PETS2007-S8-Ground-Truth-Part2.tar.gz (11MB) ]
Ground-truthing Articulated Tracking via Virtual Markers:
This is a small Matlab GUI utility to ease the tedious task of defining the ground-truth 2D (x,y) location of body joints for articulated human body tracking in videos that have no associated motion capture data. The system partially automates the task of defining the 'virtual markers' by predicting a joint's (x,y) position in the next frame via template matching, and a human can correct the prediction interactively if necessary. For each joint, it takes 5-10 minutes to ground-truth 500 frames of video. Under Windows, any installed AVI codec is supported; under Unix Matlab only supports raw video.
Download Matlab source code: [ .zip (19.2Kb) ]
Incremental clustering of dynamic data streams using connectivity based representative points:
An implementation of the RepStream incremental clustering algorithm. The source code has been compiled on Linux and Mac OS X but should compile on any platform supported by GCC. Synthetic data sets for testing are also provided.
Download source [repstream-1.0.tar.bz2 (43KB)]
[synthetic-data-sets.tar.bz2 (46KB)]
