popsift

Why we included this project

PopSift is worth a look if your pipeline already leans on scale-invariant keypoint detection and the CPU-bound OpenCV path is slowing you down. It is a CUDA implementation of SIFT that stays close to David Lowe's original paper, which matters when you want the classic behavior instead of the speed/accuracy trade-offs some faster ports make. The library exposes a clean asynchronous enqueue/get API with a config object, so you can pull keypoints, scales, and descriptors from grayscale frames on NVIDIA hardware without reworking your whole feature extraction stage. Integration is straightforward via CMake's find_package, and the included demo lets you sanity-check the output before wiring it into a larger system. Teams doing structure-from-motion, image matching, or real-time localization on NVIDIA GPUs should find it a practical fit.