OCS is an open source GPU implemenation of SIFT feature detection and matching. The CUDA code is a direct port of OpenCV's SIFT implementation.
- Added an option to disable subpixel interpolation
- Added an option to limit the maximum sample radius used in orientation assignment and descriptor computation
- Output descriptor element range is [-1,1] instead of [-128,127]
- Removed the "char rounding" at the end of descriptor computation
- Added an option to limit the maximum number of octave layers
- Added an option to upscale the image before feature detection
- Use a fixed gaussian blur kernel size of 9x9
For all meassurements the median time of 50 executions was taken. Host-Device data transfers for example uploading the image for feature detection is not included. The test code can be found in src/test.cu and is run by default.
Sift Parameters:
- doubleScale = false
- octaveLayers = 3
- contrastThresold = 0.04
- edgeThreshold = 10
- sigma = 1.6
- maxDescriptorRadius = 16
| Image Size | 640x480 | 1280x720 | 1920x1080 | 3000x1688 | 4096x2304 |
|---|---|---|---|---|---|
| #Features | 990 | 1408 | 1700 | 3382 | 6184 |
| GTX 1080 | 0.851 | 1.311 | 1.856 | 3.642 | 6.247 |
| GTX 970 | 1.917 | 2.880 | 4.165 | 8.240 | 14.133 |
Bruteforce knn search with k=4
| Match Size | 990x990 | 1408x1408 | 1700x1700 | 3382x3382 | 6184x6184 |
|---|---|---|---|---|---|
| GTX 1080 | 0.281 | 0.559 | 0.837 | 3.275 | 11.010 |
| GTX 970 | 0.708 | 1.396 | 2.095 | 7.976 | 25.953 |
| GTX 760 | 1.161 | 2.311 | 3.472 | 13.231 | 44.369 |
//load image and copy it to the device
cv::Mat1f img1 = cv::imread("data/landscape.jpg",cv::IMREAD_GRAYSCALE);
Saiga::CUDA::CudaImage<float> d_img(img.cols,img.rows,Saiga::iAlignUp(img.cols*sizeof(float),256));
copyImage(Saiga::MatToImageView<float>(img),d_img,cudaMemcpyHostToDevice);
//initialize sift and init memory.
int maxFeatures = 10000;
SIFTGPU sift(d_img.width,d_img.height,false,-1,maxFeatures,3,0.04,10,1.6);
sift.initMemory();
//extract keypoints and descriptors on the gpu
thrust::device_vector<SiftPoint> keypoints(maxFeatures);
thrust::device_vector<float> descriptors(maxFeatures * 128);
int extractedPoints = sift.compute(d_img, keypoints, descriptors)For running the sample.
The OCS ocde is licensed under the MIT License, but you also have to comply with OpenCV's license. The SIFT algorithm itself is patented and not free for commercial use.