A GPU implementation of David Lowe's Scale Invariant Feature Transform
Changchang wu
http://cs.unc.edu/~ccwu
University of North Carolina at Chapel Hill
1. SIFT
SIFTGPU is an implementation of SIFT for GPU. SiftGPU uses GPU to process pixels and features
parallely in Gaussian pyramid construction, DoG keypoint detection and descriptor generation
for SIFT. Compact feature list is efficiently build through a GPU/CPU mixed reduction.
SIFTGPU is inspired by Andrea Vedaldi's sift++ and Sudipta N Sinha et al's GPU-SIFT. Many
parameters of sift++ ( for example, number of octaves,number of DOG levels, edge threshold,
etc) are available in SiftGPU.
SIFTGPU also includes a GPU exhaustive/guided sift matcher SiftMatchGPU. It basically multiplies
the descriptor matrix on GPU and find closest feature matches on GPU. GLSL/CUDA/CG implementations
are all provided.
NEW: The latest SIFTGPU also enables you to use Multi-GPUs and GPUS on different computers.
Check doc/manual.pdf for more information. You can modify some marcros definition in
SimpleSIFT.cpp and speed.cpp to enable the testing of the new functions.
2. Requirements
The default implemntation uses GLSL, and it requires a GPU that has large memory and supports
dynamic branching. For nVidia graphic cards, you can optionally use CG(require fp40) or
CUDA implementation. You can try different implementations and to find out the fastest one
for different image sizes and parameters.
The GLSL version may not work on ATI now. They did compile sucessfully with ATI Catalyst 8.9,
but not any more with 9.x versions.
SiftGPU uses DevIl Image library, GLEW and GLUT. You'll need to make sure your system has
all the dependening libraries. SiftGPU should be able to run on any operation system that supports
the above libraries
For windows system visual studio solution are provided as msvc/SiftGPU.dsw, msvc/SiftGPU.sln and
msvc/SiftGPU_CUDA_Enabled.sln. Linux/Mac makefile is in folder Linux of the package.
3. Helps
Use -help to get parameter information. Check /doc/manual.pdf for samples and explanations.
In the vc workspace, there is a project called SimpleSIF that gives an example of simple
SiftGPU usage. There are more examples of different ways of using SiftGPU in manual.pdf
Check /doc/manual.pdf for help on the viewer.
