Vision Benchmark Suite
|Download (Version 1.3.2)|
(Note: If you are trying to download from China, and are having problems with the above link, it is because of the Great Firewall of China. Email me below to get a copy.)
The San Diego Vision Benchmark Suite (SD-VBS) is a suite of diverse vision applications drawn from the vision domain. It is intended to help architects, compiler writers, and system designers study the construction of future systems that excel at vision-oriented applications. Additionally, vision codes tend to have a fair amount of parallelism, which makes them a good candidate for formulating future multicore and parallel architectures.
The applications are drawn from the current state-of-the-art in computer vision, in consultation with vision researchers. Each benchmark is provided in both MATLAB and C form.
MATLAB is the preferred language of vision researchers, while C makes it easier to map the applications to research platforms. The C code minimizes pointer usage and employs clean constructs to make them easier for parallelization.
Furthermore, we provide a spectrum of input sets that enable researchers to control simulation time, and to understand properties as inputs increase to leverage better processor performance.
|Disparity Map||Motion, Tracking and Stereo Vision|
|Feature Tracking||Motion, Tracking and Stereo Vision|
|Image Segmentation||Image Analysis|
|Scale Invariant Feature Transform (SIFT)||Image Analysis|
|Maximally Stable Regions (MSER)||Image Analysis|
|Robot Localization||Image Understanding|
|Support Vector Machines (SVM)||Image Understanding|
|Image Stitch||Image Processing and Formation|
|Texture Synthesis||Image Processing and Formation|
Now at 1300 users!
Now at 700 users!
Now at 350 users!
|01.15.11|| U Penn has parallelized several of the benchmarks for their 2012 HPCA paper! |
Version 1.3.2 Released.
This version replaces stack allocation with a heap allocation for Image Segmentation benchmark.
Version 1.3.1 Released.
This version includes bug fixes to the Image Segmentation benchmark.
Version 1.3 Released.
This version improves memory allocation/deallocation so that SD-VBS code can be incorporated into vision applications.
|03.11.10||VBS is now being used by over a hundred researchers!|
Version 1.22 Released.
This version includes bug fixes to the timing functions.
Version 1.21 Released.
This version excludes wuxga input data set from all the benchmarks since the sizes of wuxga and fullhd were similar (1920x1200 and 1920x1080 respectively).
Repository size: 421MB
Version 1.2 Released!!
This version includes bug fixes and has more input data sets (vga, fullhd-1080p, wuxga).
Repository size: 622MB
|11.23.09||Approximate Run Times
We give the run times, in cycles, below, on a 2.6 GHz Opteron system, with a 1MB cache, for different inputs.
Version 1.2 (coming soon) of the benchmark suite will include larger inputs
(including 800x600 and 1920x1080) for those benchmarks that scale with
image size for those who are looking for larger working sets and run times.
Version 1.01 Released!!
|11.21.09||Face Detection will be released at a later date; MSER has been substituted for it. |
Prof. Brian Demsky of UC Irvine has Feature Tracking parallelized using his parallelizing compiler: 26x speedup on 62 tiles!