SiftGPU

点击打开链接

SiftGPU: A GPU Implementation of Scale Invariant Feature Transform (SIFT)

Changchang Wu

University of North Carolina at Chapel Hill

* VisualSFM with integration of SiftGPU and Multicore Bundle Adjustment
* Multicore Bundle Adjustment for GPU and CPU is now available
* Last update on 2/7/2012 (Bugfix: update maximum texture dimension for CUDA-based SiftMatchGPU). 
* 4/30/2012, Mac OSX patch to ProgramGLSL.cpp provided by Roy Shilkrot: ProgramGLSL.cpp.patch 

SIFT Implementation

SiftGPU is an implementation of SIFT [1] for GPU. SiftGPU processes pixels parallely to build Gaussian pyramids and detect DoG Keypoints. Based on GPU list generation[3], SiftGPU then uses a GPU/CPU mixed method to efficiently build compact keypoint lists. Finally keypoints are processed parallely to get their orientations and descriptors.

SiftGPU is inspired by Andrea Vedaldi's sift++[2] and Sudipta N Sinha et al's GPU-SIFT[4] . Many parameters of sift++ ( for example, number of octaves, number of DOG levels, edge threshold, etc) are also available in SiftGPU. The shader programs are dynamically generated according to the parameters that user specified.

SiftGPU also includes a GPU exhaustive/guided sift matcher SiftMatchGPU. It basically multiplies the descriptor matrix on GPU and finds the closest feature matches on GPU. Both GLSL and CUDA implementations are provided. 

Requirements

SiftGPU requires a decent GPU that has a large graphic memory and supports dynamic branching. GLSL is used by default, and CUDA is provided as an alternative for nVidia graphic cards.

SiftGPU uses GLEW 1.51, DevIL1.77 (can be disabled), GLUT(only by the viewer), and CUDA(optional). You'll need to make sure that your system has all the depending libraries of corresponding versions. To update the libraries, you'll need to replace the header files in SiftGPU\Include\, and the corresponding binaries.

NOTE FOR CUDA : 1. The thread block setting is currently tuned on nVidia GTX 8800. It may not be optimized for other GPUs. 2. The CUDA version is not compiled by default. You need to define CUDA_SIFTGPU_ENABLED to the compiler and recompile the package. For VS2010 users, you can just use SiftGPU_CUDA_Enabled solution. 

Download

SiftGPU-V382 (5.0MB; Including code, manual , windows binary and some test images) Want to cite SiftGPU? 
VisualSFM: an integrated SFM system based on SiftGPU and Multicore Bundle Adjustment
Mac OSX patch to ProgramGLSL.cpp provided by Roy Shilkrot: ProgramGLSL.cpp.patch 
You might be interested in the Matlab Versions mex'd by Adam Chapman and by Parag. K. Mital 
* For commercial licensing, please contact Trude Amick ([e-mail], 919-843-7415)

SimpleSIFT.cpp gives some examples of using SiftGPU and SiftMatchGPU. 

NOTICE: download V382 or do the follow to fix a minor bug (2/7/2012)
please add "GlobalUtil::_texMaxDimGL = 32768;" after L1373 of ProgramCU.cu 
In V370-V381, the maximum texture dimension is not properly updated from hardware. 

List of recent changes (complete list, previous versions) 
  * Really fixed the bug in CUDA-version Fundamental-matrix guided SIFT matching (12/04/2011)
  * Automatic reduction of working dimension according to GPU memory cap (8/2011)
  * Fixed the orientation GLSL function that may have problem with new nVidia driver (7/2011)
  * Modified some GLSL code for ATI and tested on a real ATI card (7/2011)
  * Converted the MSVC Solution to Visual Studio V2010 and tested CUDA4 (6/2011)
  * Automatic switching from OpenGL to CUDA when OpenGL is not supported (1/2011)
  * Dropped the indirect data transfer path CPU->GL->CUDA (1/2011) 

Some previous changes  
  6. Added device selection for Multi-threading (Check the example atMultiThreadSIFT.cpp). 
  5. Used SSE to speedup the descriptor normalization step for the OpenGL implementation.  
  4. Added CUDA-based SiftGPU/SiftMatchGPU implementation. See Figure below for the speed.  
  3. Added OpenGL-based sift matching implementation, check example #7 in manual. (Thanks to Zach) 
  2. Added function to compute descriptors for user-specified keypoints, check example #6 in manual. 
  1. Improved speed by %50 compared with V293. Look here for experiment details andexplanations

Evaluation

Below is the evaluation of the speed of V340 on different image sizes. "-fo -1" means using upsampled image. "-glsl" uses GLSL and "-cuda" uses CUDA (The experiment images are all resized from this image) .  

 System : nVidia 8800GTX, 768MB, Driver 182.08, Windows XP, Intel 3G P4 CPU, 3.5G RAM. (V311 Speed) 

Below is the comparision with Lowe's SIFT on box.pgm using the comparison code fromVedaldi's SIFT .

References

[1]   D. G. Lowe. Distinctive image features from scale-invariant keypoints . International Journal of Computer Vision, November 2004. 
[2]   A. Vedaldi. sift++, http://vision.ucla.edu/~vedaldi/code/siftpp/siftpp.html.
[3]   G. Ziegler, et al. GPU point list generation through histogram pyramids. In Technical Report, June 2006. 
[4]   Sudipta N Sinha, Jan-Michael Frahm, Marc Pollefeys and Yakup Genc, "GPU-Based Video Feature Tracking and Matching ", 
        EDGE 2006, workshop on Edge Computing Using New Commodity Architectures, Chapel Hill, May 2006