Dynamic Mean Shift: Adaptive Mode-Seeking

• Dynamic Mean Shift is a family of adaptive mode-seeking techniques that iteratively update kernel metrics, active sets, or state representations.
• It employs dynamic updates—such as adaptive target models, evolving feature metrics, and grid-based representations—to address limitations in static clustering and object tracking.
• Recent developments integrate dynamic mean shift into differentiable neural architectures, graph formulations, and interacting particle systems, broadening its practical applications.

Dynamic mean shift refers to mean-shift-based procedures in which the mode-seeking update is coupled to an evolving state rather than a fixed kernel-density geometry. Across the literature, that evolving state may be an adaptive target model and bandwidth in visual tracking, an iteration-dependent feature metric, a dynamically recomputed active neighborhood, a moving grid-cell representation, a graph rate matrix, a differentiable attention operator, an interacting-particle system with variable weights, or a diffusion-model sampling rule (Mohammadi et al., 2012, Chakraborty et al., 2020, Kumar et al., 2022, Craig et al., 2021, Belhadji et al., 14 Feb 2025, Thamizharasan et al., 21 Feb 2025). Taken together, these works suggest that dynamic mean shift is best understood as a family of adaptive mode-seeking constructions rather than a single canonical algorithm.

1. Baseline mean shift and the motivation for dynamization

Classical mean shift updates a current iterate to a kernel-weighted local average. One standard form is

$$y_{t,i} \leftarrow \frac{\sum_{j \in [n]} K\!\left(\frac{\lVert y_{t-1,i}-y_{t-1,j}\rVert}{h}\right)y_{t-1,j}}{\sum_{j \in [n]} K\!\left(\frac{\lVert y_{t-1,i}-y_{t-1,j}\rVert}{h}\right)},$$

where $h$ is the bandwidth and $K$ is the kernel (Jang et al., 2021). Equivalent formulations write the update as a gradient-ascent step on a kernel density estimator, of the form

$$x_{\mathrm{new}} \longleftarrow x + c\,h^2\,\frac{\nabla \hat p_{n,h}(x)}{\hat p_{n,h}(x)},$$

which makes explicit that mean shift is a density-gradient operator (Xiang et al., 2016).

The motivation for dynamic variants arises when the static assumptions implicit in the baseline update fail. In color-based object tracking, the traditional algorithm is described as having a “lack of dynamism in its target model,” making it unsuitable for objects with changes in their sizes and shapes; this directly motivated an adaptive-model and adaptive-bandwidth tracker (Mohammadi et al., 2012). In classifier-based tracking, a related limitation is that a single fixed information source can be unreliable across changing scenes, which motivated dynamic confidence-based fusion of two classifiers, with their contributions recalculated from frame to frame using correlations between histograms of their weight images and a defined ideal weight image in the previous frame (Topkaya et al., 2011).

A common misconception is that “dynamic mean shift” names one specific estimator. The literature does not support that reading. The recurring invariant is the mean-shift principle of moving toward higher estimated density or toward a local weighted mean; the dynamic component is the update law, the representation of locality, or the object being shifted.

2. Adaptive metrics, active sets, and accelerated state updates

One major line of work makes the kernel geometry itself iteration-dependent. In the feature-weighted blurring mean shift algorithm, the Euclidean norm is replaced by

$$\|x-y\|_w=\sqrt{\sum_{l=1}^p w_l(x_l-y_l)^2}, \qquad \mathbf 1^\top w=1,\; w_l\ge 0,$$

and the weights are updated after each shift step by a softmax over per-feature within-cluster residual sums of squares (Chakraborty et al., 2020). The method is therefore dynamic in the precise sense that the kernel metric changes from iteration to iteration through the current feature weights. Its underlying objective is an entropy-regularized weighted within-cluster sum of squares, and the paper states rigorous theoretical convergence guarantees with a convergence rate of at least a cubic order.

A second line replaces smooth reweighting by an evolving active set. For the generalized mean shift with triangular kernel profile,

$$C_n=\{\, i \mid d(\hat{x}_n,x_i)<1 \,\}, \qquad \hat{x}_{n+1} = \mu(C_n)\in \arg\min_x \sum_{i\in C_n} d(x,x_i).$$

Here the dynamic element is the repeated recomputation of the active set $C_n$ from the current iterate (Razakarivony et al., 2020). The paper proves a strong finite-termination property: the sequence is stationary regardless of the distance used and converges in a finite number of steps.

A third line changes the state representation for computational reasons. MeanShift++ replaces pointwise neighbor search by a density-weighted mean of adjacent grid cells, using the current cell and its neighboring cells in $\{-1,0,1\}^d$; the resulting runtime is linear in the number of points and exponential in dimension, making the method appropriate for low-dimensional applications such as image segmentation and object tracking (Jang et al., 2021). GridShift goes further by moving active grid cells rather than individual data points. Active cells shift, can merge when they collapse to the same location, and update their neighborhoods after each shift; the per-iteration cost is $\mathcal{O}(m3^d)$, where $m$ is the number of active grid cells (Kumar et al., 2022). In both cases, the mode-seeking logic is preserved, but the dynamic object is no longer the original point cloud alone.

Adaptive execution strategies form a related computational variant. A GPU-accelerated faster mean shift for Euclidean data uses Seed Selection and Early Stopping, with CPU-side adjustment of the number of seeds based on the estimated cluster count. The paper characterizes this as an acceleration of conventional mean shift rather than a new objective, and reports around 3 times speedup compared to state-of-the-art GPU-based mean-shift algorithms with optimized GPU memory consumption on a 200K-point clustering problem (You et al., 2021).

3. Tracking, segmentation, and biomedical motion estimation

Dynamic mean shift first became especially visible in tracking, where both the target appearance and the search geometry evolve over time. A colored object tracker with adaptive model and bandwidth was proposed specifically to address the inability of traditional color-based mean shift to handle changes in object size and shape; the proposed method is described as a fast novel threephase colored object tracker algorithm based on the mean shift idea while utilizing adaptive model, with feasible, robust, and acceptable speed relative to other algorithms (Mohammadi et al., 2012).

Confidence-based dynamic classifier combination extends mean-shift tracking through adaptive fusion of information sources. The method uses two different classifiers, one coming from a background modeling method, to generate a weight image. Their contributions are calculated dynamically using their confidences, so that the final weight image changes with the estimated reliability of each classifier (Topkaya et al., 2011). This modification keeps the mean-shift search mechanism but replaces a fixed appearance model by a confidence-weighted ensemble.

In medical imaging, tumor motion tracking in liver ultrasound uses mean shift as the displacement estimator between consecutive frames. The target in the first frame is defined using an ellipse; edge, texture, and intensity features are extracted from the first frame, and mean shift is applied to each feature separately to find the center of the ellipse related to that feature in the next frame (Rasekhi, 2015). The center used for motion estimation is the weighted average of these centers. Once the correct ellipsoid in each frame is known, the method applies the Dynamic Directional Gradient Vector Flow version of active contour models to find the correct tumor boundary, and translated boundary sample points serve as the initial guess for the next frame’s active contour. The reported interpretation is that mean shift estimates target movement, while the contour stage refines boundary geometry.

Grid-based accelerations also produced tracking systems. MeanShift++ proposes object tracking using precomputed grid cells instead of explicit histograms, and reports promising results for object tracking (Jang et al., 2021). GridShift similarly introduces a new object-tracking algorithm based on active grid-cell clusters rather than a back-projected color histogram, with a tracking window that adapts as the object appears larger or smaller and can be enlarged when no matching points are found (Kumar et al., 2022). This suggests a broader shift in dynamic mean shift tracking: the adaptive component is no longer limited to bandwidth selection, but extends to the representation of appearance, the search window, and the state propagated across frames.

4. Dynamical-systems views, denoising operators, and convergence theory

A substantial theoretical literature treats mean shift not merely as a pointwise heuristic but as a dynamical system acting on densities. One PDE-based analysis models dynamic clustering by a conservation law in which the local velocity is proportional to the density gradient divided by the density. In the unsupervised case, this yields the anti-diffusion equation

$h$0

and the paper concludes that unsupervised mean-shift type algorithms are intrinsically unstable (Wang et al., 2012). Its central claim is especially restrictive: correct convergence for the unsupervised algorithm is possible only when the original probability density is transformed into a multivariate normal distribution with no dependence structure. The same work argues that a more stable and convergent mean shift algorithm might be achieved by adopting a judiciously chosen supervision mechanism, introduced through an added sink term.

A different theoretical program interprets mean shift as a distribution operator. The generalized operator

$h$1

induces a transformed distribution $h$2, allowing analysis of how probability mass is relocated after one mean-shift step (Xiang et al., 2016). Under smoothness, Morse, and boundary-gradient assumptions, this operator increases probability mass inside high-density level sets by order $h$3, increases density at modes by order $h$4, and decreases density at local minima by order $h$5. Repeated application compounds the concentration effect, which explains the use of mean shift as a denoising preprocessor for clustering, two-sample testing, and anomaly detection.

Convergence results for the iterates themselves have become more general. A recent analysis frames the mean shift update as a majorization-minimization step and proves that, under fairly mild conditions involving differentiability, Lipschitz continuity of the gradient, and the Łojasiewicz property, the mode estimate sequence has finite length and converges to a critical point of the kernel density estimate (Yamasaki et al., 2023). The same paper evaluates the convergence rate through the Łojasiewicz exponent: finite-time convergence if $h$6, exponential or linear convergence if $h$7, and polynomial convergence if $h$8. It is notable for covering the biweight kernel, which earlier analytic-kernel results did not cover.

Taken jointly, these analyses show that dynamic mean shift has two distinct theoretical faces. One face is anti-diffusive and potentially unstable when interpreted as unsupervised density evolution. The other face is a convergent ascent procedure under explicit kernel and regularity assumptions. The coexistence of these perspectives explains why dynamic modifications are often introduced together with adaptive supervision, regularization, or restricted kernel classes.

5. Graph, hyperspherical, and differentiable formulations

Dynamic mean shift has also been reformulated on non-Euclidean state spaces. On graphs, mean shift can be represented by a rate matrix that moves probability mass only along edges toward neighbors with higher potential. This graph mean shift can then be interpolated with graph diffusion by

$h$9

which yields the continuum Fokker–Planck interpolation

$K$0

The parameter $K$1 therefore governs a balance between density-driven ascent and geometry-driven diffusion (Craig et al., 2021). The same work shows that an extended diffusion-map family indexed by $K$2 realizes the same interpolation, and that the limit $K$3 approaches a graph version of KNF mean shift. This reframes dynamic mean shift as a continuum of clustering dynamics rather than a single update equation.

A neural-network formulation appears in the Mean Shift Mask Transformer for unseen object instance segmentation. The method simulates the von Mises-Fisher mean shift clustering algorithm by using object queries on a hypersphere and a hypersphere attention mechanism,

$K$4

with $K$5 (Lu et al., 2022). Each decoder layer simulates one iteration of the clustering, and multiple decoder layers replace the mean shift clustering process. The model is therefore dynamic in a literal architectural sense: cluster centers are iteratively refined by learned, differentiable mean-shift analogues, with a fixed number of decoder layers standing in for a fixed number of iterations. The reported motivation is that traditional mean shift clustering is not differentiable and therefore difficult to integrate into end-to-end training, whereas this construction allows joint training and inference of feature extraction and clustering.

These formulations broaden the notion of locality. In graph settings, locality is edge-constrained transport and Fokker–Planck drift-diffusion. In hyperspherical neural settings, locality is cosine-like affinity on the unit sphere together with masked query refinement. Both preserve the mean-shift idea of repeated local mode seeking, but they relocate it into different ambient geometries.

6. Interacting particles, diffusion distillation, and the current interpretation of the field

Recent work has extended dynamic mean shift into generative modeling and quantization. Mean-shift distillation introduces a diffusion distillation technique in which the mean-shift vector

$K$6

serves as a proxy for the gradient of the diffusion output distribution (Thamizharasan et al., 21 Feb 2025). The construction starts from the classical fixed-point identity for Gaussian-kernel mean shift and replaces unavailable samples from the target density by samples from a product distribution that combines the learned diffusion density with a Gaussian kernel centered at the current iterate. In practice, the update is the displacement $K$7, where $K$8 is sampled from that product distribution. The paper states that the extrema are aligned with the modes, that the method is a drop-in replacement for score distillation sampling requiring neither model retraining nor extensive modification of the sampling procedure, and that it exhibits superior mode alignment as well as improved convergence in synthetic and practical setups.

A distinct but related development arises in MMD-optimal quantization. Weighted quantization using MMD derives a Wasserstein–Fisher–Rao gradient flow for the MMD energy and discretizes it as an interacting-particle system with both position and weight dynamics (Belhadji et al., 14 Feb 2025). The resulting fixed-point algorithm, mean shift interacting particles (MSIP), extends classical mean shift by moving from one particle to many interacting particles with variable weights. When $K$9, the construction recovers the classical mean-shift fixed-point equation; for general $$x_{\mathrm{new}} \longleftarrow x + c\,h^2\,\frac{\nabla \hat p_{n,h}(x)}{\hat p_{n,h}(x)},$$0, the paper shows that MSIP can be interpreted as preconditioned gradient descent and as a relaxation of Lloyd’s algorithm. Its contribution is therefore not merely another heuristic generalization, but a unification of gradient flows, mean shift, and MMD-optimal quantization.

These contemporary formulations also clarify the limits of the term. Some dynamic variants are principally adaptive approximations or accelerations rather than new objectives; MeanShift++ is explicitly framed as a fast, grid-based approximation to standard MeanShift (Jang et al., 2021), and GPU-accelerated faster mean shift is explicitly framed as an accelerated implementation strategy (You et al., 2021). Some have theoretical guarantees only for a component of the method; MeanShift++ provides guarantees for the grid-based density estimator but not a full convergence analysis of the complete iterative clustering algorithm (Jang et al., 2021). For MSIP, a full convergence theory is described as still open (Belhadji et al., 14 Feb 2025).

Accordingly, dynamic mean shift occupies a broad methodological space. It may denote adaptive visual tracking, feature-weighted clustering, active-set mode seeking, grid-collapsed acceleration, graph drift-diffusion interpolation, differentiable hyperspherical attention, interacting-particle quantization, or mode-seeking diffusion distillation. This suggests that the unifying object is not a unique algorithmic template, but the preservation of mean shift’s local mode-seeking principle under an evolving state representation.