FFD: Fast Feature Detector

Scale-invariance, good localization and robustness to noise distortions are the main properties that a local feature detector should possess. Most existing detectors find excessive unstable points increase number of keypoints be matched computational time matching step. In this paper, we show robust accurate exist in specific scale-space domain. To end, first formulate superimposition problem into mathematical model then derive closed-form solution for multiscale analysis. The is formulated via difference-of-Gaussian (DoG) kernels continuous domain, it proved setting pyramid's blurring ratio smoothness 2 0.627, respectively, facilitates detection reliable keypoints. For applicability proposed discrete images, discretize using undecimated wavelet transform cubic spline function. Theoretically, complexity our method less than 5\% popular baseline Scale Invariant Feature Transform (SIFT). Extensive experimental results superiority over representative hand-crafted learning-based techniques accuracy time. code supplementary materials can found at~{\url{https://github.com/mogvision/FFD}}.