Another CUDA SIFT

An(other) cu(DA) Sift is a CUDA implementation (among many: PopSift, cuSIFT, etc.) of the Scale Invariant Feature Transform (SIFT) algorithm that detects and describes local features in images.

This implementation closely follows the version by opencv, structured largely through the same flow, achieving real-time results without going through extensive optimizations and keeping the code rather readable.

I started out this repository as a project for me to learn CUDA, I'm also rather bad at template programming. If you see design choices that are poop or are generally considered bad, it's probably because it is.

Performance

On an RTX 4070S, matching between 2 sets of 2000 128D descriptor takes just under 1ms. Detect and Compute (not counting memory transfer time):

752x480	1920x1200	1600x900
0.95ms	3.1ms	2.5ms
84MiB	298MiB	214 MiB

Dependencies

Hardware: This repo is developed on an RTX 4070 Super, the code compiled and succesfully run for sm_61. I currently do not have the capacity to test the code on other hardwares.

Software:

0. bazel for build. g++10, CUDA version 12.6.
1. rules_cuda for cuda toolchains with bazel
2. zlib and msgpack-c for serializing and saving intermediate states of the algorithm.
3. cli11 for command line argument parsing.

Item 1-3 should be automatically handled by bazel.

4. OpenCV, as I used opencv's image utility functions to load the images and draw matches. To use the examples provided in tool, opencv is required.

Installation

To build OpenCV dependency, do the following:

# In thirdparty/opencv/build run the following line.
# It should install everything into thirdparty/opencv_bin
$ cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=../../opencv_bin ..
# Install opencv, specify the number of cores to use if you don't want your computer to freeze up 
$ make install -j

Use as third party package

# In MODULE.bazel
bazel_dep(name = "another-cuda-sift", version = "")
git_override(
    module_name = "another-cuda-sift",
    remote = "https://github.com/Cryptum169/another-cuda-sift.git",
    commit = "commit-sha" #
)

# In BUILD
cc_binary(
    name = "user_program",
    srcs = ["user_program.cc"],
    deps = ["@another-cuda-sift//sift_cuda/interface:interface"],
)

Usage

tool/extract_and_match_example.cc provided an example usage. The file here loads a list of images under a directory, detects and computes the descriptor for keypoints on each image, matches keypoints between consecutive images, and plotted the results.

CudaSiftConfig config;
config.col_width = r;
config.row_width = c;
sift_cuda::Detector detector(config);
detector.gpuWarmUpAndAllocate();

for (...) {
    detector.detectAndCompute(host_img[idx]);
    detector.copyToHost( /* descriptor = */ false);
}

Known Issues and FAQ

See faq.md

License

SIFT's patent has expired on 2020-03-06. See Google Patents