The estimation of the gradient of a density function, with applications in pattern recognition

By Keinosuke Fukunaga, Larry D. Hostetler. 1975

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Benchmark and Survey of Automated Machine Learning Frameworks (2019, methodology): “For each classification algorithm, all configuration vectors are aggregated using mean shift clustering (Fukunaga and Hostetler, 1975) with a bandwidth h = 0.”
Survey on Automated Machine Learning (2019, methodology): “e decided to only compare congurations with each other having the same classication algorithm. For each classication algorithm, all conguration vectors are aggregated using mean shift clustering (Fukunaga and Hostetler, 1975) with a bandwidth h= 0:25. To account for the mixed-type vector representat”
Learning in Implicit Generative Models (2016, methodology): “There are a number of tools that ar e available though, including the methodof-moments [18], the empirical likelihood [30], Monte Carl o sampling [25], and mean-shift estimation [14].”
Understanding Convolutional Neural Networks for Text Classification (2018, methodology): “We use Mean Shift Clustering (Fukunaga and Hostetler, 1975; Cheng, 1995), an algorithm that does not require specifying an a-priori number of clusters, and does not make assumptions about their shapes.”
Semantic Instance Segmentation with a Discriminative Loss Function (2017, methodology): “We display the images after 0, 2, 4, 8, 16, 32 and 64 gradient update steps. issue, we make the clustering more robust against outliers by applying a fast variant of the mean-shift algorithm [14].”
Recurrent Pixel Embedding for Instance Grouping (2017, methodology): “Mean-shift and closely related algorithms [26, 13, 14, 15] use kernel density estimation to approximate the probability density from a set of samples and then perform clustering on the input data by assigning or moving each sample to the nearest mode (local maxima).”
V3D enables real-time 3D visualization and quantitative analysis of large-scale biological image data sets (2010, methodology): “In the 3D editing mode (Supplementary Video 9), V3D-Neuron automatically represents the entire neuron model as the aggregation of segments, each of which is bounded by a pair of tip or branching nodes.”
Regularized Multitask Learning for Multidimensional Log-Density Gradient Estimation (2014, methodology): “1 Mode-seeking clustering A practical application of log-density gradient estimation is mode-seeking clustering (Fukunaga and Hostetler, 1975; Cheng, 1995; Comaniciu and Meer, 2002; Sasaki et al., 2014).”
Quick Shift and Kernel Methods for Mode Seeking (2008, background): “Mean shift [9, 5] is based on an efficient rule to evolve the trajectories yi(t) when the window k(x) can be written as ψ(‖x‖(2)2) for a convex function ψ(z) (for instance the Gaussian window has ψ(z) ∝ exp(−z)).”
Improved fast gauss transform and efficient kernel density estimation (2003, methodology): “Specifically the mean shift algorithm [11, 4, 5] is chosen as a case study for the IFGT.”
Pattern Recognition (2015, unknown): “After the forward propagation of Deeplab-VGG16 backbone network [29] and the extra layers (see in Fig.”
Neural Empirical Bayes (2019, background): “Mode seeking is a deep topic in the kernel density literature around the mean shift algorithm (Fukunaga and Hostetler, 1975; Cheng, 1995; Comaniciu and Meer, 2002), also around the intriguing topic of “ghost modes” (Carreira-Perpi˜n´an and Williams, 2003).”
A Novel Ranking-Based Clustering Approach for Hyperspectral Band Selection (2016, methodology): “Meanwhile, the FDPC approach only requires measuring the distance between all pairs of data points and does not need to parameterize a probability distribution [34] or a multidimensional density function [35], which is much easier to be computed.”
Wild Variational Approximations (2016, background): “In this case a density estimator q̂, e.g. kernel density estimators (KDE) [31] or those parametrised by a neural network [32], is fitted to the samples {zk = f( k,x)} ∼ q, and the gradient of log q is approximated as ∇φ log q(z|x) ≈ ∇z log q̂(z|x)∇φf .”