/CudaSift

A CUDA implementation of SIFT for NVidia GPUs (1.2 ms on a GTX 1060)

Primary LanguageCudaMIT LicenseMIT

CudaSift - SIFT features with CUDA

This is the fourth version of a SIFT (Scale Invariant Feature Transform) implementation using CUDA for GPUs from NVidia. The first version is from 2007 and GPUs have evolved since then. This version is slightly more precise and considerably faster than the previous versions and has been optimized for Kepler and later generations of GPUs.

On a GTX 1060 GPU the code takes about 1.2 ms on a 1280x960 pixel image and 1.7 ms on a 1920x1080 pixel image. There is also code for brute-force matching of features that takes about 2.2 ms for two sets of around 1900 SIFT features each.

The code relies on CMake for compilation and OpenCV for image containers. OpenCV can however be quite easily changed to something else. The code can be relatively hard to read, given the way things have been parallelized for maximum speed.

The code is free to use for non-commercial applications. If you use the code for research, please cite to the following paper.

M. Björkman, N. Bergström and D. Kragic, "Detecting, segmenting and tracking unknown objects using multi-label MRF inference", CVIU, 118, pp. 111-127, January 2014. ScienceDirect

Update in feature matching (2019-05-17)

The brute force feature matcher has been significantly improved in speed. The largest improvements can be seen for large feature sets with 10000 features or more, but as can be seen below, it performs rather well even with just 2000 features. The file match.pdf includes a description of the optimizations done in this version.

New version for Pascal (2018-10-26)

There is a new version optimized for Pascal cards, but it should work also on many older cards. Since it includes some bug fixes that changes slightly how features are extracted, which might affect matching to features extracted using an older version, the changes are kept in a new branch (Pascal). The fixes include a small change in ScaleDown that corrects an odd behaviour for images with heights not divisible by 2^(#octaves). The second change is a correction of an improper shift of (0.5,0.5) pixels, when pixel values were read from the image to create a descriptor.

Then there are some improvements in terms of speed, especially in the Laplace function, that detects DoG features, and the LowPass function, that is seen as preprocessing and is not included in the benchmarking below. Maybe surprisingly, even if optimizations were done with respect to Pascal cards, these improvements were even better for older cards. The changes involve trying to make each CUDA thread have more work to do, using fewer thread blocks. For typical images of today, there will be enough blocks to feed the streaming multiprocessors anyway.

Latest result of version under test:

1280x960 1920x1080 GFLOPS Bandwidth Matching
Turing GeForce RTX 2080 Ti 0.42* 0.56* 11750 616 0.30*
Pascal GeForce GTX 1080 Ti 0.58* 0.80* 10609 484 0.42*
Pascal GeForce GTX 1060 1.2 1.7 3855 192 2.2
Maxwell GeForce GTX 970 1.3 1.8 3494 224 2.5
Kepler Tesla K40c 2.4 3.4 4291 288 4.7

Matching is done between two sets of 1911 and 2086 features respectively. A star indicates results from the last checked in version.

Benchmarking of new version (2018-08-22)

About every 2nd year, I try to update the code to gain even more speed through further optimization. Here are some results for a new version of the code. Improvements in speed have primarilly been gained by reducing communication between host and device, better balancing the load on caches, shared and global memory, and increasing the workload of each thread block.

1280x960 1920x1080 GFLOPS Bandwidth Matching
Pascal GeForce GTX 1080 Ti 0.7 1.0 10609 484 1.0
Pascal GeForce GTX 1060 1.6 2.4 3855 192 2.2
Maxwell GeForce GTX 970 1.9 2.8 3494 224 2.5
Kepler Tesla K40c 3.1 4.7 4291 288 4.7
Kepler GeForce GTX TITAN 2.9 4.3 4500 288 4.5

Matching is done between two sets of 1818 and 1978 features respectively.

It's questionable whether further optimization really makes sense, given that the cost of just transfering an 1920x1080 pixel image to the device takes about 1.4 ms on a GTX 1080 Ti. Even if the brute force feature matcher is not much faster than earlier versions, it does not have the same O(N^2) temporary memory overhead, which is preferable if there are many features.

Benchmarking of previous version (2017-05-24)

Computational cost (in milliseconds) on different GPUs:

1280x960 1920x1080 GFLOPS Bandwidth Matching
Pascal GeForce GTX 1080 Ti 1.7 2.3 10609 484 1.4
Pascal GeForce GTX 1060 2.7 4.0 3855 192 2.6
Maxwell GeForce GTX 970 3.8 5.6 3494 224 2.8
Kepler Tesla K40c 5.4 8.0 4291 288 5.5
Kepler GeForce GTX TITAN 4.4 6.6 4500 288 4.6

Matching is done between two sets of 1616 and 1769 features respectively.

The improvements in this version involved a slight adaptation for Pascal, changing from textures to global memory (mostly through L2) in the most costly function LaplaceMulti. The medium-end card GTX 1060 is impressive indeed.

Usage

There are two different containers for storing data on the host and on the device; SiftData for SIFT features and CudaImage for images. Since memory allocation on GPUs is slow, it's usually preferable to preallocate a sufficient amount of memory using InitSiftData(), in particular if SIFT features are extracted from a continuous stream of video camera images. On repeated calls ExtractSift() will reuse memory previously allocated.

#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <cudaImage.h>
#include <cudaSift.h>

/* Reserve memory space for a whole bunch of SIFT features. */
SiftData siftData;
InitSiftData(siftData, 25000, true, true);

/* Read image using OpenCV and convert to floating point. */
cv::Mat limg;
cv::imread("image.png", 0).convertTo(limg, CV32FC1);
/* Allocate 1280x960 pixel image with device side pitch of 1280 floats. */ 
/* Memory on host side already allocated by OpenCV is reused.           */
CudaImage img;
img.Allocate(1280, 960, 1280, false, NULL, (float*) limg.data);
/* Download image from host to device */
img.Download();

int numOctaves = 5;    /* Number of octaves in Gaussian pyramid */
float initBlur = 1.0f; /* Amount of initial Gaussian blurring in standard deviations */
float thresh = 3.5f;   /* Threshold on difference of Gaussians for feature pruning */
float minScale = 0.0f; /* Minimum acceptable scale to remove fine-scale features */
bool upScale = false;  /* Whether to upscale image before extraction */
/* Extract SIFT features */
ExtractSift(siftData, img, numOctaves, initBlur, thresh, minScale, upScale);
...
/* Free space allocated from SIFT features */
FreeSiftData(siftData);

Parameter setting

The requirements on number and quality of features vary from application to application. Some applications benefit from a smaller number of high quality features, while others require as many features as possible. More distinct features with higher DoG (difference of Gaussians) responses tend to be of higher quality and are easier to match between multiple views. With the parameter thresh a threshold can be set on the minimum DoG to prune features of less quality.

In many cases the most fine-scale features are of little use, especially when noise conditions are severe or when features are matched between very different views. In such cases the most fine-scale features can be pruned by setting minScale to the minimum acceptable feature scale, where 1.0 corresponds to the original image scale without upscaling. As a consequence of pruning the computational cost can also be reduced.

To increase the number of SIFT features, but also increase the computational cost, the original image can be automatically upscaled to double the size using the upScale parameter, in accordance to Lowe's recommendations. One should keep in mind though that by doing so the fraction of features that can be matched tend to go down, even if the total number of extracted features increases significantly. If it's enough to instead reduce the thresh parameter to get more features, that is often a better alternative.

Results without upscaling (upScale=False) of 1280x960 pixel input image.

thresh #Matches %Matches Cost (ms)
1.0 4236 40.4% 5.8
1.5 3491 42.5% 5.2
2.0 2720 43.2% 4.7
2.5 2121 44.4% 4.2
3.0 1627 45.8% 3.9
3.5 1189 46.2% 3.6
4.0 881 48.5% 3.3

Results with upscaling (upScale=True) of 1280x960 pixel input image.

thresh #Matches %Matches Cost (ms)
2.0 4502 34.9% 13.2
2.5 3389 35.9% 11.2
3.0 2529 37.1% 10.6
3.5 1841 38.3% 9.9
4.0 1331 39.8% 9.5
4.5 954 42.2% 9.3
5.0 611 39.3% 9.1