betaFeedback

WorksPhysics and Astronomy

Particle Image Velocimetry Refinement via Consensus ADMM for Active Fluid Control

Abstract

Particle Image Velocimetry (PIV) is among the central modalities for measuring flow fields across laboratory, industrial and environmental setting. Traditional PIV approaches typically depend on tuning parameters specific to the imaging setup, making the performance sensitive to variations in illumination, flow conditions, and seeding density. Similarly, state-of-the-art machine learning methods for flow quantification are fragile outside their training set. In our experiments, we observed that flow quantification would improve if different tunings (or algorithms) were applied to different regions of the same image pair. Motivated by this observation, we thus pose flow quantification as a multi-estimator fusion problem: several heterogeneous algorithms process the same image pair in parallel, and their dense flow fields are treated as complementary estimates. To fuse them, we adopt a consensus framework based on the alternating direction method of multipliers, incorporating priors such as smoothness and incompressibility. We perform several numerical experiments to demonstrate the benefits of this approach. For instance, we achieve a decrease in end-point-error of up to 20% of a dense-inverse-search estimator at an inference rate of 60Hz, and we show how performance can be increased with outlier rejection. Our method is implemented in JAX and integrated into Flow Gym, enabling reproducible comparisons with the state of the art and systematic evaluation across different base algorithms. Finally, we demonstrate successful deployment of our method in the same real-world active-fluids-control setup of Terpin and D'Andrea [1], where a reinforcement-learning agent uses our flow estimates to learn to minimize drag (down by 36%) or maximize it (up to 32%) with only two minutes of real-world interaction. Hardware and software are made available at ActiveFluidControl.com.

Introduction

Particle Image Velocimetry (PIV) [2, 3, 4] and related image-based flow-quantification methods [5, 6, 7] are classical tools for measuring flow fields across experimental fluid mechanics, with applications spanning laboratory [2], industrial [3], microfluidic [8], [9], and environmental settings [10], [11]. Traditionally, their role has been primarily observational: to reconstruct velocity fields for analysis, validation, and diagnostics. For this reason, performance has classically been assessed mainly through accuracy, often using the end-point error (EPE), i.e., the average ℓ2\ell_2 norm of the difference between the ground truth and the estimated value of the flow field.

Recent results in active fluid control motivate a broader perspective. In their recent study, Terpin and D’Andrea [1] showed that, in a real-world fluids-control task, an agent equipped with dense flow feedback could discover high-performance strategies within minutes of real-world interaction, whereas removing that feedback caused learning to fail altogether. This observation introduces a different axis in the evaluation of PIV algorithms. Once flow estimation is embedded in a real-time learning or control loop, latency becomes equally central. A highly accurate but slow estimator cannot run online, whereas a fast estimator with poor accuracy may provide feedback that is too crude to support learning. In this regime, the relevant objective is the speed–accuracy trade-off. This tension is already visible in current methods: using Flow Gym [12], one can verify that the state-of-the-art RAFT32-PIV [13] reaches only up to 1.7 Hz on an RTX 4090 GPU, while fast estimators such as DIS [14] can operate at 60 Hz or more, but at substantially lower accuracy.

There are, broadly speaking, two ways to improve this

Corresponding author: Antonio Terpin ([email protected]). operating point: one may attempt to accelerate highly accurate but computationally expensive estimators, or instead improve the accuracy of methods that already satisfy the real-time constraint. In this work, we pursue the latter direction. A central obstacle is that no single estimator or parameter setting performs uniformly best across all flows, or even across all regions of the same image pair. In practice, classical flow estimators often require parameter choices that depend on the experimental setup [15], and recent work has begun to explore dynamic parameter adaptation to mitigate this sensitivity [16]. From our experience, it is often easier to tune an estimator for a specific regime than to identify a single configuration that performs well across heterogeneous conditions. This effect can arise even within a single flow field. In Figure 2, we compare two parameter tunings of DIS on the same PIV image pair and observe that each tuning is more accurate on different spatial regions. This suggests that robustness may be improved not by searching for a single universal estimator, but by combining complementary candidate estimates and enforcing global consistency across the resulting field.

Flow estimates and agreement map for two PIV parameter tunings

Figure 2. On the left and middle, we report the flow estimates (colors represent the magnitude of the flow field at each pixel) for the same PIV image pair of two different tunings of the JAX DIS implementation from Flow Gym [12]. On the right, we mark in green pixels in which both estimates have an estimate closer than 0.1px to the ground-truth, light blue if only the first tuning, orange if only the second, and black if both estimates fail to be closer than 0.1px to the ground-truth.

Contributions. Motivated by this observation, we pose flow quantification as a multi-estimator fusion problem: several heterogeneous algorithms process the same image pair, and their dense velocity fields are treated as complementary observations. To the best of our knowledge, this is the first formulation and experimental demonstration of parallel estimator-output fusion for fluid velocimetry. We combine the different estimates via consensus Alternating Direction Method of Multipliers (ADMM), enforcing priors such as smoothness and incompressibility; Figure 1. In this paper, we showcase the results of this approach using as base algorithms DIS [14], DeepFlow [17], and Farnebäck [18], but the method applies to any other algorithm, possibly learning-based. For instance, with DIS, we achieve a decrease in EPE of up to 20 % at an inference rate of 60 Hz, and we show how performance can be increased with outlier-rejection. Our method is implemented in JAX and integrated into Flow Gym [12], enabling reproducible evaluation and direct use in downstream control tasks. In particular, we deploy it on the real-world active-fluids-control setup of [1], where a reinforcement-learning agent uses our flow estimates to learn to minimize drag (down by 36 %) or maximize it (up to 32 %) with only two minutes of real-world interaction.

The proposed consensus-ADMM pipeline for PIV refinement

Figure 1. The proposed consensus-ADMM pipeline for PIV refinement. Consecutive PIV images of illuminated tracer particles are processed by multiple estimators in parallel, and their outputs are fused through consensus ADMM into a refined flow field.

Notation: We let nn be the number of pixels. The single-channel input frames are I0,I1I_0,I_1. Since Δt\Delta t and the pixel-to-world conversion are known, estimating velocity is equivalent to estimating the pixel displacement (ux,uy)∈R2(u_x,u_y) \in\mathbb{R}^2 at each pixel. We denote the full flow field by u=(ux,uy)∈R2n\mathbf{u}=(u_x,u_y)\in\mathbb{R}^{2n}, with entries indexed by (γ)ℓ(\gamma)_\ell, ℓ∈{1,…,2n}\ell\in\{1,\ldots,2n\}.

PIV refinement

In PIV, flow estimation is commonly formulated as the minimization of an image-consistency term together with a regularizer. Under the usual small-displacement assumption, the short inter-frame delay often makes brightness constancy a reasonable approximation, i.e., I0(x,y)≈I1(x+ux,y+uy)I_0(x,y)\approx I_1(x+u_x,y+u_y). This leads to objectives of the form

min⁡u∈R2nJd(u,(I0,I1))+λJr(u,(I0,I1)),(1)\min_{\mathbf{u}\in\mathbb{R}^{2n}} \mathcal{J}_d(\mathbf{u},(I_0,I_1))+\lambda\mathcal{J}_r(\mathbf{u},(I_0,I_1)), \tag*{(1)}

where Jd\mathcal{J}_d measures consistency with the image pair and Jr\mathcal{J}_r encodes priors such as smoothness.

In our setting, however, we are not interested in solving (1) from scratch with a single estimator. Instead, given a collection of PIV algorithms A={A1,…,AN}\mathcal{A}=\{A_1,\ldots,A_N\} producing candidate flows u^i=Ai(I0,I1)∈R2n\hat{\mathbf{u}}_i=A_i(I_0,I_1)\in\mathbb{R}^{2n}, we seek a refined field that remains close to these candidates where they are reliable, while enforcing global priors. This leads to the optimization problem

min⁡u∈R2n∑i=1NJi(u,Ai(I0,I1),(I0,I1))+g(u,(I0,I1)),(2)\min_{\mathbf{u}\in\mathbb{R}^{2n}} \sum_{i=1}^{N}\mathcal{J}_i(\mathbf{u},A_i(I_0,I_1),(I_0,I_1))+g(\mathbf{u},(I_0,I_1)), \tag*{(2)}

where each fif_i measures agreement with the ii-th candidate estimate, and gg collects global regularization terms expressing priors such as smoothness and incompressibility.

PIV refinement via Consensus ADMM

In (2), each data term fi(u,(I0,I1))f_i(\mathbf{u},(I_0,I_1)) is defined by the individual algorithm estimate u^i=Ai(I0,I1)\hat{\mathbf{u}}_i=A_i(I_0,I_1) and by the image pair (I0,I1)(I_0,I_1). By equivalently writing (2) as

min⁡u1,…,uN,u∈R2n∑i=1Nfi(ui,(I0,I1))+g(u,(I0,I1))s.t.ui−u=0,i=1,…,N(3)\begin{aligned} \min_{\mathbf{u}_1,\ldots,\mathbf{u}_N,\mathbf{u}\in\mathbb{R}^{2n}}\quad&\sum_{i=1}^{N} f_i(\mathbf{u}_i,(I_0,I_1))+g(\mathbf{u},(I_0,I_1))\\ \text{s.t.}\quad&\mathbf{u}_i-\mathbf{u}=0,\quad i=1,\ldots,N \tag*{(3)} \end{aligned}
ui−u=0,i=1,…,N(4)\mathbf{u}_i-\mathbf{u}=0,\quad i=1,\ldots,N \tag*{(4)}

we match the formulation in [19, Problem 7.2]. The coupling constraints ui=u\mathbf{u}_i=\mathbf{u}, i∈{1,…,N}i\in\{1,\ldots,N\} enforce that all local estimates ui\mathbf{u}_i agree on a single consensus field u\mathbf{u}. With the dual variables di∈R2n\mathbf{d}_i\in\mathbb{R}^{2n}, the augmented Lagrangian reads

Lρ(u1,…,uN,d1,…,dN,u)=∑i=1Nfi(ui,(I0,I1))+g(u,(I0,I1))+ρ2∑i=1N∥ui−u+di∥22,\mathcal{L}_{\rho}(\mathbf{u}_1,\ldots,\mathbf{u}_N,\mathbf{d}_1,\ldots,\mathbf{d}_N,\mathbf{u})=\sum_{i=1}^{N}f_i(\mathbf{u}_i,(I_0,I_1))+g(\mathbf{u},(I_0,I_1))+\frac{\rho}{2}\sum_{i=1}^{N}\lVert\mathbf{u}_i-\mathbf{u}+\mathbf{d}_i\rVert_2^2,

where ρ>0\rho>0 is the penalty parameter; in our experiments, we use ρ=2\rho=2.

Remark (Penalty-parameter sensitivity). ADMM performance can be sensitive to the penalty parameter ρ\rho. Under standard convergence assumptions [19 Section 3.2], ρ\rho does not change the limiting solution, but it can affect the convergence rate: small values may weakly enforce consensus, whereas large values may overemphasize the proximal term relative to data fidelity. Since the relative scales of fif_i and gg may vary across heterogeneous flow regimes, the adaptive selection of ρ\rho is a promising direction for future work.

By applying the global consensus variant of ADMM [19 Section 7.2] we obtain the fixed-point iteration2:

ui[k+1]=argmin⁡ui∈R2n fi(ui,(I0,I1))+ρ2∥ui−u[k]+di[k]∥22,(5)\begin{aligned} \mathbf{u}_i[k+1] &= \underset{\mathbf{u}_i\in\mathbb{R}^{2n}}{\operatorname{argmin}}\ f_i(\mathbf{u}_i,(I_0,I_1))+\frac{\rho}{2}\lVert\mathbf{u}_i-\mathbf{u}[k]+\mathbf{d}_i[k]\rVert_2^2, \tag*{(5)} \end{aligned}
u[k+1]=argmin⁡u∈R2n g(u,(I0,I1))+ρ2∑i=1N∥u−ui[k+1]−di[k]∥22,(6)\begin{aligned} \mathbf{u}[k+1] &= \underset{\mathbf{u}\in\mathbb{R}^{2n}}{\operatorname{argmin}}\ g(\mathbf{u},(I_0,I_1))+\frac{\rho}{2}\sum_{i=1}^{N}\lVert\mathbf{u}-\mathbf{u}_i[k+1]-\mathbf{d}_i[k]\rVert_2^2, \tag*{(6)} \end{aligned}
di[k+1]=di[k]+ui[k+1]−u[k+1],(7)\mathbf{d}_i[k+1]=\mathbf{d}_i[k]+\mathbf{u}_i[k+1]-\mathbf{u}[k+1], \tag*{(7)}

where the updates (4a) and (4c) are for all i∈{1,…,N}i\in\{1,\ldots,N\}. In the remainder of this section, we show that for fi(⋅)f_i(\cdot) and g(⋅)g(\cdot) of interest for the PIV refinement problem in (2), we can efficiently compute the iterations (4a)-(4b), enabling the implementation of Algorithm 1:

Algorithm 1. PIV Refinement via Consensus ADMM

Inputs: u^i∈R2n ∀i∈{1,…,N}\hat{\mathbf{u}}_i\in\mathbb{R}^{2n}\ \forall i\in\{1,\ldots,N\}.

Initialization: u[0]=1N∑i=1Nu^i\mathbf{u}[0]=\frac{1}{N}\sum_{i=1}^{N}\hat{\mathbf{u}}_i

ui[0]=u^i ∀i∈{1,…,N}\mathbf{u}_i[0]=\hat{\mathbf{u}}_i\ \forall i\in\{1,\ldots,N\}

di[0]=u^i−u[0] ∀i∈{1,…,N}\mathbf{d}_i[0]=\hat{\mathbf{u}}_i-\mathbf{u}[0]\ \forall i\in\{1,\ldots,N\}

For k=0k=0 to K1−1K_1-1:

    Update ui[k+1]\mathbf{u}_i[k+1] according to (6) in Section II-B.

    uˉ[k+1]←1N∑i=1Nui[k+1]\bar{\mathbf{u}}[k+1]\leftarrow\frac{1}{N}\sum_{i=1}^{N}\mathbf{u}_i[k+1]

    dˉ[k]←1N∑i=1Ndi[k]\bar{\mathbf{d}}[k]\leftarrow\frac{1}{N}\sum_{i=1}^{N}\mathbf{d}_i[k]

    u[k+1]←Algorithm2⁡(uˉ[k+1],dˉ[k],u[k])\mathbf{u}[k+1]\leftarrow\operatorname{Algorithm 2}(\bar{\mathbf{u}}[k+1],\bar{\mathbf{d}}[k],\mathbf{u}[k])

    di[k+1]←di[k]+ui[k+1]−u[k+1]\mathbf{d}_i[k+1]\leftarrow\mathbf{d}_i[k]+\mathbf{u}_i[k+1]-\mathbf{u}[k+1]

Output: u[k+1]\mathbf{u}[k+1]

Standard ADMM convergence results apply to the exact convex updates in Algorithm 1 (see, e.g., [19 Section 3.2]). In the real-time implementation considered here, we instead use a fixed iteration budget of K1=30K_1=30. A systematic analysis of the sensitivity of performance to K1K_1 is omitted for brevity, and we believe that future work may address the applicability of the method at higher frame rates or on lower-power hardware, as well as how performance can improve with more computational resources.

Data term and local update.

An effective choice for fi(⋅)f_i(\cdot) is given, for a closed, proper, and convex scalar ϕ(⋅)\phi(\cdot) by

fi(ui,(I0,I1))=∑ℓ=12nϕ(wi,ℓ(I0,I1)(ui−u^i)ℓ),(8)f_i(\mathbf{u}_i,(I_0,I_1))=\sum_{\ell=1}^{2n}\phi\left(w_{i,\ell}(I_0,I_1)(\mathbf{u}_i-\hat{\mathbf{u}}_i)_\ell\right), \tag*{(8)}

where wi,ℓ(I0,I1)≥0w_{i,\ell}(I_0,I_1)\ge0 encodes the estimate and pixel-wise confidence in u^i\hat{\mathbf{u}}_i and may depend on the image pair (I0,I1)(I_0,I_1). Hereafter, we will drop the dependency on I0,I1I_0,I_1 for ease of notation in all the quantities depending on I0,I1I_0,I_1. Then, the update (4a) boils down to solving

argmin⁡ui∈R2n ∑ℓ=12nϕ(wi,ℓ(ui−u^i)ℓ)+ρ2∑ℓ=12n(ui−u[k]+di[k])ℓ2.\underset{\mathbf{u}_i\in\mathbb{R}^{2n}}{\operatorname{argmin}}\ \sum_{\ell=1}^{2n}\phi\left(w_{i,\ell}(\mathbf{u}_i-\hat{\mathbf{u}}_i)_\ell\right)+\frac{\rho}{2}\sum_{\ell=1}^{2n}\left(\mathbf{u}_i-\mathbf{u}[k]+\mathbf{d}_i[k]\right)_\ell^2.

Thus, with c=(u[k]−di[k])ℓ\mathbf{c}=(\mathbf{u}[k]-\mathbf{d}_i[k])_\ell and r=(u[k]−u^i−di[k])ℓ\mathbf{r}=(\mathbf{u}[k]-\hat{\mathbf{u}}_i-\mathbf{d}_i[k])_\ell, the update reduces to 2nN2nN scalar problems:

argmin⁡t∈R ϕ(wi,ℓ(t−(u^i)ℓ))+ρ2(t−c)2= (u^i)ℓ+argmin⁡t∈R(1ρϕ(wi,ℓt)+12(t−r)2).(9)\underset{t\in\mathbb{R}}{\operatorname{argmin}}\ \phi\left(w_{i,\ell}(t-(\hat{\mathbf{u}}_i)_\ell)\right)+\frac{\rho}{2}(t-c)^2 =\ (\hat{\mathbf{u}}_i)_\ell+\underset{t\in\mathbb{R}}{\operatorname{argmin}}\left(\frac{1}{\rho}\phi(w_{i,\ell}t)+\frac{1}{2}(t-r)^2\right). \tag*{(9)}

That is, the update (4a) amounts to a per-pixel proximal evaluation. For many ϕ\phi of interest, this can be done in closed form. For instance, this is the case for the ℓ1\ell_1 norm, the squared ℓ2\ell_2 norm, and the Huber loss; cf. Appendix A. We compare these alternatives in Section III.

The weights wi,ℓw_{i,\ell} can be computed in several different ways. In Section III, we show that wi,ℓ=1w_{i,\ell}=1 performs well, but we also consider different weighting schemes; cf. also Appendix B. Finally, if the estimates are subject to an outlier-rejection scheme before being provided to our method, one can simply set the weights corresponding to outliers to zero and (4a)-(4c) will implicitly perform data interpolation according to the other estimates and priors.

Regularization term and global update.

We embed some of the priors on the solution via g(⋅)g(\cdot):

g(u)=λsRs(u)+λcurvRcurv(u)+λdivRdiv(u),(10)g(\mathbf{u})=\lambda_s\mathcal{R}_s(\mathbf{u})+\lambda_{\mathrm{curv}}\mathcal{R}_{\mathrm{curv}}(\mathbf{u})+\lambda_{\mathrm{div}}\mathcal{R}_{\mathrm{div}}(\mathbf{u}), \tag*{(10)}

where each λ∙≥0\lambda_{\bullet}\ge0 is a tunable parameter. Each penalty encodes a prior about the spatial regularity or structure of the flow field:

  • Smoothness: We penalize sharp variations across estimates of nearby pixels,

Rs(u)=∥Dxux∥22+∥Dyux∥22+∥Dxuy∥22+∥Dyuy∥22,(11)\mathcal{R}_s(\mathbf{u})=\lVert D_xu_x\rVert_2^2+\lVert D_yu_x\rVert_2^2+\lVert D_xu_y\rVert_2^2+\lVert D_yu_y\rVert_2^2, \tag*{(11)}

where Dx,DyD_x,D_y are the first-order central finite-difference operators along the xx- and yy-directions, evaluated only on the interior grid; accordingly, derivative-based quantities are computed on the cropped domain obtained by excluding one-pixel boundaries. This penalty term suppresses high spatial frequencies and is straightforward to discretize, but may oversmooth motion discontinuities.

  • Second-order spatial variation: We penalize second-order spatial variations of the flow by convolving (∗)(*) each velocity component with a standard discrete 3×33 \times3 Laplacian stencil (equivalently, a narrow-scale Laplacian-of-Gaussian (LoG) approximation [20]):

Rcurv(u)=∥LoG∗ux∥22+∥LoG∗uy∥22.(12)\mathcal{R}_{\mathrm{curv}}(\mathbf{u}) = \|\mathrm{LoG} * u_x\|_2^2 + \|\mathrm{LoG} * u_y\|_2^2. \tag*{(12)}
  • Incompressibility: We penalize non-zero values of the divergence of the flow field:

Rdiv(u)=∥∇⋅u∥22=∥Dxux+Dyuy∥22.(13)\mathcal{R}_{\mathrm{div}}(\mathbf{u}) = \|\nabla\cdot\mathbf{u}\|_2^2 = \|D_xu_x + D_yu_y\|_2^2. \tag*{(13)}

The parameters λ∙\lambda_{\bullet} are easy to tune: they represent how much the optimization should weigh certain regularity priors. We describe a simple empirical procedure in Appendix C.

With this choice, the update in (4b) is a strictly convex quadratic problem, which we solve approximately by running K2=30K_2 = 30 iterations of gradient descent with the Adam optimizer [21], using default hyperparameters and step size η=0.01\eta= 0.01. The overall procedure is summarized in Algorithm 2. This design integrates naturally with our JAX pipeline and yields a simple fixed-budget implementation compatible with the real-time setting targeted in this work. At the same time, the optimality condition of (4b) defines a sparse structured linear system induced by the finite-difference and convolution stencils appearing in the regularization terms. Exploiting this structure more directly through matrix-free Krylov [22], multigrid methods [23], conjugate gradient iterations [24], and hard-constrained neural networks [25] is a promising future research direction.

Algorithm 2. Global update

Choose: η\eta

Inputs: uˉ[k+1],dˉ[k],u[k]\bar{\mathbf{u}}[k+1], \bar{\mathbf{d}}[k], \mathbf{u}[k]

Initialization: z[0]=u[k]\mathbf{z}[0] = \mathbf{u}[k]

For j=0j = 0 to K2−1K_2 - 1:

∇z[j]L(z[j])=∇z[j](g(z[j])+Nρ2∥z[j]−uˉ[k+1]−dˉ[k]∥22)\nabla_{\mathbf{z}[j]}\mathcal{L}(\mathbf{z}[j]) = \nabla_{\mathbf{z}[j]}\left(g(\mathbf{z}[j]) + \frac{N\rho}{2}\|\mathbf{z}[j] - \bar{\mathbf{u}}[k+1] - \bar{\mathbf{d}}[k]\|_2^2\right)

z[j+1]←Adam⁡step⁡(z[j],∇z[j]L(z[j]),η)\mathbf{z}[j+1] \leftarrow\operatorname{Adam}\operatorname{step}(\mathbf{z}[j], \nabla_{\mathbf{z}[j]}\mathcal{L}(\mathbf{z}[j]), \eta)

Output: z[K2]\mathbf{z}[K_2]

Numerical Results

In this section, we numerically assess the performance of our method. The code to reproduce all the experiments is available in Flow Gym [12]. We consider the widely adopted PIV dataset [13], which comprises 256×256256 \times256 image pairs and ground-truth flow data, utilizing SynthPix as the dataloader [26]. For all numerical experiments, we evaluate the proposed method using three base estimators available in Flow Gym [12]: DIS [14], DeepFlow [17], and Farnebäck [18]. For DIS and Farnebäck, we construct three parameter configurations by tuning each method on 10%10\% of the training split of [13], with one configuration tuned on each of the following class groups: (Cylinder, JHTDB), (Backstep, Uniform), and (SQG, DNS). We denote the resulting variants by DIS1, DIS2, DIS3 and F1, F2, F3, respectively. We further denote by DIS* and F* the best-performing DIS and Farnebäck configurations when evaluated across all categories. Since the current OpenCV implementation [27] does not expose tunable parameters for DeepFlow, we use a single DeepFlow instance, denoted by DF. We apply the proposed framework to the estimator collections A∈{{DIS1,DIS2,DIS3},{F1,F2,F3},{DIS∗,DF,F∗}}\mathcal{A} \in\{\{\mathrm{DIS1},\mathrm{DIS2},\mathrm{DIS3}\},\{\mathrm{F1},\mathrm{F2},\mathrm{F3}\},\{\mathrm{DIS}^*,\mathrm{DF},\mathrm{F}^*\}\}. Our goal is to quantify the benefit of the consensus refinement itself, independently of the underlying base estimators. Accordingly, for each collection A\mathcal{A}, we compare the refined output against the best individual estimator in A\mathcal{A} on the test split of [13], which we denote by A∗\mathcal{A}^*. Since we are not interested in comparisons between different A\mathcal{A}, we measure performance by reporting the relative change in dataset-average EPE with respect to this baseline. Let AEPE⁡(A;D)\operatorname{AEPE}(\mathcal{A};\mathcal{D}) be the average EPE attained by the algorithm on the dataset D\mathcal{D}. Then, we define the rAEPE(A,A∗)(\mathcal{A}, \mathcal{A}^*) as

AEPE⁡(A;D)−AEPE⁡(A∗;D)AEPE⁡(A∗;D)×100 [%].(14)\frac{\operatorname{AEPE}(\mathcal{A};\mathcal{D}) - \operatorname{AEPE}(\mathcal{A}^*;\mathcal{D})}{\operatorname{AEPE}(\mathcal{A}^*;\mathcal{D})} \times100\ [\%]. \tag*{(14)}

Negative values of rAEPE therefore indicate an improvement over the baseline A∗\mathcal{A}^*.

For the ablation studies, we use a separate 10%10\% split of the train set from [13], which we refer to as the validation set, and we tune the regularization parameters on an additional 10%10\% split of the train set. Often, PIV involves outlier rejection and data interpolation schemes [28]. To investigate the dependency of our framework on the quality of the outlier-rejection scheme and to assess its intrinsic data interpolation properties, we define, for τ≥0\tau\ge0, wi,ℓTw^T_{i,\ell} as wi,ℓT=wi,ℓw^T_{i,\ell} = w_{i,\ell} if (EPE⁡i)ρ(ℓ)≤τ(\operatorname{EPE}_{i})_{\rho(\ell)} \le\tau, where ρ(ℓ)\rho(\ell) is the pixel and entry corresponding to ℓ\ell, and wi,ℓT=0w^T_{i,\ell} = 0 otherwise. We use the oracle outlier-rejection protocol as a controlled diagnostic experiment to decouple the evaluation of the consensus layer from any specific practical rejection heuristic. This setting reveals the performance attainable under progressively cleaner local estimates. As we show below, the results make clear that the proposed method remains effective even in the absence of any outlier-rejection mechanism. In Section III-E, we also consider a simple learning-based rejection module and show that it can recover the oracle’s performance. Because the development of a high-performance rejection scheme is not the main scope of this work, this experiment is intended mainly as a proof of feasibility beyond the oracle setting.

Benefits of consensus

In this section, we assess the benefits of the proposed method across different baseline algorithms sets A\mathcal{A}. For reference, we also report results obtained by directly aggregating the different estimates using pixel-wise median and average.

Experimental setup. For all A\mathcal{A}, we report the rAEPE [%][\%] (11) on the test set for varying outlier-rejection threshold τ\tau. We consider ϕ(⋅)\phi(\cdot) in fi(⋅)f_i(\cdot) to be the Huber loss; see Appendix A. We tune the regularization parameters individually for each A\mathcal{A} as reported in Appendix C, and we set the weights wi,ℓw_{i,\ell} as described in Appendix § 7.2.2.

Results. We summarize the results in Figure 3. Notably, for all baseline algorithms A\mathcal{A}, our consensus method yields a considerable performance improvement across all outlier-rejection thresholds. For instance, Farnebäck tops the 40% improvement with outlier rejection, and DIS the 20% across all τ\tau. Overall, these results suggest that a simple consensus layer on top of existing estimators may substantially improve their performance, and one can choose the estimators to balance accuracy and inference speed. Importantly, ADMM-based consensus also consistently outperforms simpler aggregation baselines such as pixel-wise mean and median.

Figure 3. Benefits of consensus across different sets of algorithms A\mathcal{A}; see Section III-A. We report the rAEPE [%] (11) (lower is better) over the outlier-rejection threshold τ\tau [px].

The effects of different data terms

An issue with EPE as a metric, and consequently the consensus objective in § 2-B is that regions of fast flow (corresponding to vectors with higher magnitude) are given more importance than slow regions. Thus, different data-terms fi(⋅)f_i(\cdot) may be more or less robust to different flows. We perform this ablation in this section.

Experimental setup. We fix A={DIS1,DIS2,DIS3}\mathcal{A} = \{\mathrm{DIS}_1,\mathrm{DIS}_2,\mathrm{DIS}_3\}, and for all the options of fi(⋅)f_i(\cdot) described in Appendix A (ℓ1\ell_1 norm, squared ℓ2\ell_2 norm, Huber loss) we report the rAEPE [%] (11) on the validation set for varying outlier-rejection threshold τ\tau. We tune the regularization parameters individually for each configuration as reported in Appendix C, and we set the weights wi,ℓw_{i,\ell} as described in Appendix B.2. For the ℓ1\ell_1 loss, we use squared confidence weights. This reparameterization renders numerically comparable to the ℓ2\ell_2 and Huber variants the influence of the weights in the proximal update.

Results. We report the results in Figure 4. The Huber loss ranks as the most reliable: it consistently improves over the baseline across thresholds and degrades the least when τ\tau is large, thanks to its built-in robustness to outliers. The ℓ1\ell_1 norm is somewhat robust but can perform poorly when τ\tau is very small (too few inliers), and the squared ℓ2\ell_2 norm can work well only when outliers are already tightly filtered; without good rejection, it gets heavily corrupted by outliers and can make the consensus worse than the baseline.

Figure 4. Impact of data terms fi(⋅)f_i(\cdot); see Section III-B. We report the rAEPE [%] (11) (lower is better) over the outlier-rejection threshold τ\tau [px].

Weighting strategy

In this section, we assess the impact of the different weighting strategies in the data term (5).

Experimental setup. We fix A={DIS1,DIS2,DIS3}\mathcal{A} = \{\mathrm{DIS}_1,\mathrm{DIS}_2,\mathrm{DIS}_3\} and compare different weighting strategies:

  • Uniform averaging: i.e., wi,ℓ=1w_{i,\ell}=1.

  • PE weighting: inverse photometric error as a confidence measure; cf. Appendix B.1.

  • (Gradient-)Adjusted PE weighting: inverse photometric error adjusted based on the image texture; cf. Appendix B.2.

  • (Gradient-)Adjusted uniform averaging: uniform confidence over all algorithms adjusted based on the image texture; cf. Appendix B.3.

In all cases, we consider the Huber loss (cf. Appendix A) and we tune the regularization parameters individually for each configuration as reported in Appendix C. We report the rAEPE [%] (11) on the validation set for varying outlier-rejection threshold τ\tau.

Results. We summarize the results in Figure 5. The results highlight the benefits of the gradient-adjusted PE weighting scheme over the naive PE weighting, as well as the perhaps surprising efficacy of simple uniform averaging.

Figure 5. Impact of weighting schemes; see Section III-C. We report the rAEPE [%] (11) (lower is better) over the outlier-rejection threshold τ\tau [px].

Effects of regularization

In this section, we assess the effects of regularization (7).

Experimental setup. We fix A={DIS1,DIS2,DIS3}\mathcal{A} = \{\mathrm{DIS}_1,\mathrm{DIS}_2,\mathrm{DIS}_3\} and consider ϕ(⋅)\phi(\cdot) in fi(⋅)f_i(\cdot) to be the Huber loss; see Appendix A. To understand the effects of regularization independently of the other elements of the pipeline, we also consider a

fictitious weighting strategy in which wi,ℓ=1w_{i,\ell}=1 if and only if algorithm AiA_i provided the best estimate at the corresponding location, i.e.,

wi,ℓ={1if i=argmin⁡j∈{1,2,3} ∥u∗(xℓ,yℓ)−u^j(xℓ,yℓ)∥22,0otherwise,(15)w_{i,\ell} = \begin{cases} 1 & \text{if } i = \underset{j\in\{1,2,3\}}{\operatorname{argmin}}\ \lVert\mathbf{u}^{*}(x_\ell,y_\ell)-\hat{\mathbf{u}}_j(x_\ell,y_\ell)\rVert_2^2,\\ 0 & \text{otherwise,} \end{cases} \tag*{(15)}

where u∗\mathbf{u}^{*} is the ground-truth, (xℓ,yℓ)(x_\ell,y_\ell) is the pixel corresponding to the ℓ\ell-th entry, and we consider an arbitrary tiebreaking strategy so that the argmin is a singleton. Overall, we compare four methods:

  • wi,ℓw_{i,\ell} as in Appendix B.2, without regularization (λ∙=0\lambda_{\bullet}=0).

  • wi,ℓw_{i,\ell} as in Appendix B.2, with regularization.

  • wi,ℓw_{i,\ell} as in (12), without regularization (λ∙=0\lambda_{\bullet}=0).

  • wi,ℓw_{i,\ell} as in (12), with regularization.

When using regularization, we select λ∙\lambda_{\bullet} as in Appendix C. Results. We summarize the results in Figure 6. Overall, all the results consistently show the benefit of regularization, as well as the benefit of better weighting schemes: The ideal weighting suggests the presence of additional performance gains, motivating future work to investigate different, and possibly learning-based, schemes.

Figure 6. Impact of the regularization; see Section III-D. We report the rAEPE [%] (11) (lower is better) over the outlier-rejection threshold τ\tau [px].

Learning-based outlier rejection

Next, we move beyond the oracle protocol and show that the proposed consensus framework can be combined with a practical learning-based outlier-rejection module.

Experimental setup. We consider the DIS and Farnebäck estimator families separately. For each family, we train a U-Net [29] to predict, at each pixel, if the local estimate has endpoint error larger than 1 pixel. At inference time, these predictions are used to reject unreliable estimates in the consensus procedure. We compare three consensus configurations: without outlier rejection, with oracle rejection at threshold τ=1\tau=1, and with the learned rejection module. Results. We report the results in Table I. The learning-based rejection module nearly recovers the gain obtained with oracle rejection at τ=1\tau=1, showing that a practical scheme can capture most of the benefit of oracle filtering.

{DIS1, DIS2, DIS3}{F1, F2, F3}
W/o rejection19.631.8253
Oracle (τ=1\tau=1)27.2443.623
U-net26.8943.568

Table I. Baseline and consensus performance with no rejection, oracle rejection (τ=1\tau=1), and learning-based rejection; see Section III-E. We report rAEPE [%] (11).

Active fluid control

We conclude by assessing if the proposed method can be deployed in a real active-fluids-control loop.

Experimental setup. We use exactly the same active-fluids-control setup, tasks, observations, rewards, and training protocol as in [1]. In particular, we consider both drag minimization and drag maximization on the spinning-cylinder system, and we repeat the training three times for each task. The only change is the flow-quantification module used to construct the flow-based observation provided to the agent: we use the proposed consensus ADMM. Within this framework, the local sub-estimators are given by the same DIS estimator as in [1], together with variants using smaller and larger patch sizes and different regularization settings. Results. We report the corresponding training curves in Figure 7 for both drag minimization and drag maximization. In both cases, the agent can be trained successfully using flow observations generated by the proposed method, achieving drag reductions of 36 % and drag increases of up to 32 % with only two minutes of real-world interaction. These results show that the consensus-based refinement is compatible with closed-loop operation in the real system and can be deployed successfully to learn effective active-fluids-control policies.

Figure 7. Training curves for the active-fluids-control tasks of drag maximization and drag minimization. The blue curve reports performance on the maximization task (left axis), and the orange curve reports performance on the minimization task (right axis). Solid and dashed lines show the mean across runs, and the shaded bands indicate the run-to-run range. The horizontal zero lines denote the no-control baseline.

ACKNOWLEDGMENTS

The authors are grateful to Florian Dörfler for providing feedback on a preliminary version of this work, and Raff D’Andrea for the infrastructure, funding and guidance. This work was supported by a ETH RobotX research grant funded through the ETH Zurich Foundation.

References

Appendix

Choices of the data term

In this subsection, we derive the reformulations of (5) when (5) is the ℓ1\ell_1, ℓ2\ell_2 norm or the Huber loss [30].

ℓ1\ell_1 norm. The proximal operator of the ℓ1\ell_1 norm is a soft-thresholding and, thus, (6) becomes:

(u^i)ℓ+sign⁡(r)max⁡(∣r∣−wi,ℓερ−1,0).(\hat{u}_i)_\ell+ \operatorname{sign}(r)\max\left(|r|-w_{i,\ell}\varepsilon\rho^{-1},0\right).

Squared ℓ2\ell_2 norm. For the squared ℓ2\ell_2 norm, (6) becomes:

2wi,ℓu^i+ρ(u[k]−di[k])(2wi,ℓ+ρ)−1.2w_{i,\ell}\hat{u}_i+\rho\left(u[k]-d_i[k]\right)\left(2w_{i,\ell}+\rho\right)^{-1}.

Huber loss. For δ≥0\delta\ge0, the Huber loss ϕδ(t)\phi_\delta(t), defined as

ϕδ(t)={12t2if ∣t∣≤δ andδ(∣t∣−12δ)otherwise,\phi_\delta(t)= \begin{cases} \frac{1}{2}t^2 & \text{if } |t|\le\delta\text{ and}\\ \delta\left(|t|-\frac{1}{2}\delta\right) & \text{otherwise}, \end{cases}

admits the proximal operator [31] for λ≥0\lambda\ge0:

prox⁡λϕδ(v)={v(1+λ)−1,if ∣v∣≤(1+λ)δ,v−λδ,if v>(1+λ)δ,v+λδ,if v<−(1+λ)δ.\operatorname{prox}_{\lambda\phi_\delta}(v)= \begin{cases} v(1+\lambda)^{-1}, & \text{if } |v|\le(1+\lambda)\delta,\\ v-\lambda\delta, & \text{if } v>(1+\lambda)\delta,\\ v+\lambda\delta, & \text{if } v<-(1+\lambda)\delta. \end{cases}

In our experiments, we use δ=1\delta=1. The update in (6) is then computed with v=wi,ℓrv=w_{i,\ell}r and λ=1/ρ\lambda=1/\rho.

Confidence weighting

Here, we discuss three notions of per-pixel confidence.

Photometric error

In optical flow literature [14, 32], a common proxy for local estimation quality is the photometric error. The photometric error measures the average residual of the brightness constancy equation, (I0(x,y)−I1(x+u^i,x(x,y),y+u^i,y(x,y)))2(I_0(x,y)-I_1(x+\hat{u}_{i,x}(x,y),y+\hat{u}_{i,y}(x,y)))^2, in a patch P(xˉ,yˉ)P_{(\bar{x},\bar{y})} centered at pixel (xˉ,yˉ)(\bar{x},\bar{y}) corresponding to the entry ℓ\ell (note that there are two values of ℓ\ell corresponding to the same (x,y)(x,y) in our notation) and (u^i,x(x,y),u^i,y(x,y))(\hat{u}_{i,x}(x,y),\hat{u}_{i,y}(x,y)) is the estimate of algorithm ii at the pixel (x,y)(x,y). We then define

wi,ℓ=PE⁡i(xˉ,yˉ)−1.(16)w_{i,\ell}=\sqrt{\operatorname{PE}_i(\bar{x},\bar{y})}^{-1}. \tag*{(16)}

This metric captures how well an estimate satisfies brightness constancy, but it does not reflect the distinctiveness of the region: in areas with low image gradient magnitude, many different displacements can produce similarly low residuals.

Gradient-adjusted photometric error

Photometric error alone does not capture local identifiability: in low-texture regions, many displacements can yield similarly small residuals. Motivated by this observation, we define

wi,ℓ=∥∇I0(x,y)∥2PE⁡i(x,y).(17)w_{i,\ell}=\frac{\lVert\nabla I_0(x,y)\rVert_2}{\sqrt{\operatorname{PE}_i(x,y)}}. \tag*{(17)}

where (x,y)(x,y) is the pixel corresponding to the entry ℓ\ell. Intuitively, the expression in (17) increases confidence where the estimator achieves a low photometric error and the starting image has a high gradient magnitude. On the other hand, the confidence decreases where the fit to the brightness constancy equation is poor or the original image has low features. This practical weighting rule is heuristic. Its scaling is motivated by an idealized variance calculation whose assumptions need not hold for the PIV algorithms used here. Regularization, coarse-to-fine processing, estimator bias in shear layers, anisotropic covariance, and higher-order effects may therefore limit its validity. Accordingly, the calculation is neither a calibrated uncertainty model nor a performance guarantee; addressing these effects remains future work.

Assumption 1. Let u^i(x,y),u∗(x,y)∈R2\hat{\mathbf{u}}_i(x,y),\mathbf{u}^{*}(x,y)\in\mathbb{R}^2 be the estimate of the ii-th algorithm and the ground truth, respectively, evaluated at pixel (x,y)(x,y). We assume that:

A1.1 The brightness constancy assumption holds, and is thus satisfied by the ground-truth flow u∗\mathbf{u}^{*}.

A1.2 The expected value E[u^i(x,y)]\mathbb{E}[\hat{\mathbf{u}}_i(x,y)] of the estimates is unbiased for all ii, (x,y)(x,y); i.e., E[u^i(x,y)]=u∗(x,y)\mathbb{E}[\hat{\mathbf{u}}_i(x,y)]=\mathbf{u}^{*}(x,y).

A1.3 The covariance matrix of the estimates satisfies Var⁡[u^i(x,y)]=σi2(x,y)I2\operatorname{Var}[\hat{\mathbf{u}}_i(x,y)]=\sigma_i^2(x,y)I_2 for all ii, (x,y)(x,y).

We can derive an upper bound on σi2(x,y)\sigma_i^2(x,y):

Proposition 1. Let u^i(x,y),u∗(x,y)∈R2\hat{\mathbf{u}}_i(x,y),\mathbf{u}^{*}(x,y)\in\mathbb{R}^2 be the estimate of the ii-th algorithm and the ground truth, respectively, evaluated at pixel (x,y)(x,y). If Assumption 1 holds and ∇I0(x,y)≠0\nabla I_0(x,y)\ne0, then the variance of the pixel estimates σi2(x,y)\sigma_i^2(x,y) (cf. A1.3) is upper bounded as

σi2(x,y)≤E[e2+2C∥g∥∥Δ∥3]∥g∥−2,(18)\sigma_i^2(x,y)\leq\mathbb{E}\left[e^2+2C\lVert g\rVert\lVert\Delta\rVert^3\right]\lVert g\rVert^{-2}, \tag*{(18)}

where g=∇I0(x,y)g=\nabla I_0(x,y), Δ=Δu^i(x,y)\Delta=\Delta\hat{\mathbf{u}}_i(x,y), C≥0C\geq0, and e=I0(x,y)−I1(x+u^i,x(x,y),y+u^i,y(x,y))e=I_0(x,y)-I_1(x+\hat{u}_{i,x}(x,y),y+\hat{u}_{i,y}(x,y)).

Proof. Each algorithm AiA_i produces an approximate solution u^i(x,y)\hat{\mathbf{u}}_i(x,y) with residual ee. By A1.1, using the inverse compositional formulation [33] yields I0(x−u^i,x(x,y),y−u^i,y(x,y))−I1(x,y)=eI_0(x-\hat{u}_{i,x}(x,y),y-\hat{u}_{i,y}(x,y))-I_1(x,y)=e and I0(x−ux∗(x,y),y−uy∗(x,y))−I1(x,y)=0I_0(x-u^*_{x}(x,y),y-u^*_{y}(x,y))-I_1(x,y)=0. Subtracting these identities gives I0(x−ux∗(x,y),y−uy∗(x,y))−I0(x−u^i,x(x,y),y−u^i,y(x,y))=−eI_0(x-u^*_{x}(x,y),y-u^*_{y}(x,y))-I_0(x-\hat{u}_{i,x}(x,y),y-\hat{u}_{i,y}(x,y))=-e. Linearizing I0I_0 around (x,y)(x,y), we obtain gTΔ=−e+rg^\mathsf{T}\Delta=-e+r, with ∣r∣≤C∥Δ∥2\lvert r\rvert\leq C\lVert\Delta\rVert^2 for some C≥0C\geq0. Hence, e2=(gTΔ)2−2rgTΔ+r2≥(gTΔ)2−2∣r∣∥g∥∥Δ∥e^2=(g^\mathsf{T}\Delta)^2-2rg^\mathsf{T}\Delta+r^2\geq(g^\mathsf{T}\Delta)^2-2\lvert r\rvert\lVert g\rVert\lVert\Delta\rVert. Using ∣r∣≤C∥Δ∥2\lvert r\rvert\leq C\lVert\Delta\rVert^2 and ∣gTΔ∣≤∥g∥∥Δ∥\lvert g^\mathsf{T}\Delta\rvert\leq\lVert g\rVert\lVert\Delta\rVert, we get e2≥(gTΔ)2−2C∥g∥∥Δ∥3e^2\geq(g^\mathsf{T}\Delta)^2-2C\lVert g\rVert\lVert\Delta\rVert^3. Taking expectations yields E[(gTΔ)2]≤E[e2]+2C∥g∥E[∥Δ∥3]\mathbb{E}[(g^\mathsf{T}\Delta)^2]\leq\mathbb{E}[e^2]+2C\lVert g\rVert\mathbb{E}[\lVert\Delta\rVert^3]. For constant u∗(x,y)\mathbf{u}^{*}(x,y), A1.2 gives E[Δ]=0\mathbb{E}[\Delta]=0. Therefore, E[(gTΔ)2]=gTVar⁡[Δ]g\mathbb{E}[(g^\mathsf{T}\Delta)^2]=g^\mathsf{T}\operatorname{Var}[\Delta]g. Moreover, for constant u∗(x,y)\mathbf{u}^{*}(x,y), I0(x,y)I_0(x,y), and I1(x,y)I_1(x,y), we have Var⁡[Δ]=Var⁡[u^i(x,y)]\operatorname{Var}[\Delta]=\operatorname{Var}[\hat{\mathbf{u}}_i(x,y)]. Thus, by A1.3, gTVar⁡[Δ]g=σi2(x,y)∥g∥2g^\mathsf{T}\operatorname{Var}[\Delta]g=\sigma_i^2(x,y)\lVert g\rVert^2. Therefore, σi2(x,y)∥g∥2≤E[e2]+2C∥g∥E[∥Δ∥3]\sigma_i^2(x,y)\lVert g\rVert^2\leq\mathbb{E}[e^2]+2C\lVert g\rVert\mathbb{E}[\lVert\Delta\rVert^3]. We conclude dividing by ∥g∥2\lVert g\rVert^2. □

:::

The expected value E[ei2]\mathbb{E}[e_i^2] can be computed using the one-sample estimate: E[ei2]≈ei2=PE⁡i(x,y)\mathbb{E}[e_i^2]\approx e_i^2=\operatorname{PE}_i(x,y). Motivated by Proposition 1, and under the heuristic assumption that the higher-order term is small, we approximate

wi,ℓ=1σi≈∥∇I0(x,y)∥2ei≈∥∇I0(x,y)∥2PE⁡i(x,y).(19)w_{i,\ell}=\frac{1}{\sigma_i}\approx\frac{\lVert\nabla I_0(x,y)\rVert_2}{e_i}\approx\frac{\lVert\nabla I_0(x,y)\rVert_2}{\sqrt{\operatorname{PE}_i(x,y)}}. \tag*{(19)}

Gradient-adjusted uniform weighting

When λ∙=0\lambda_{\bullet}=0, (16) and (19) yield the same consensus variables given the same initial estimates, as there is no coupling across pixels. For this reason, inspired by this similarity, we also consider a third additional weighting scheme that adjusts the uniform scheme with gradient information wi,ℓ=∥∇I0(x,y)∥2w_{i,\ell}=\lVert\nabla I_0(x,y)\rVert_2.

Regularization hyperparameters

Without any aim of optimal hyperparameter tuning, we tune λ∙\lambda_{\bullet} by starting with λ∙=0\lambda_{\bullet}=0 and increasing, in order and as long as the performances on the validation set (we use 10 % of the train set in [13]) improve, λdiv,λcurv,λs\lambda_{\mathrm{div}},\lambda_{\mathrm{curv}},\lambda_s:

A={DIS1,DIS2,DIS3}\mathcal{A}=\{\mathrm{DIS1},\mathrm{DIS2},\mathrm{DIS3}\}, Huberλs\lambda_sλcurv\lambda_{\mathrm{curv}}λdiv\lambda_{\mathrm{div}}
wi,ℓ=1w_{i,\ell}=105300
wi,ℓw_{i,\ell} as in Appendix B.100.21
wi,ℓw_{i,\ell} as in Appendix B.23301300
wi,ℓw_{i,\ell} as in Appendix B.3200550750
wi,ℓw_{i,\ell} as in (12)0110
A={DIS1,DIS2,DIS3}\mathcal{A}=\{\mathrm{DIS1},\mathrm{DIS2},\mathrm{DIS3}\}, wi,ℓw_{i,\ell} as in Appendix B.2
ϕ(⋅)\phi(\cdot)λs\lambda_sλcurv\lambda_{\mathrm{curv}}λdiv\lambda_{\mathrm{div}}
ℓ1\ell_11502501100
ℓ2\ell_255001000
Huber3301300
Huber, wi,ℓw_{i,\ell} as in Appendix B.2
A\mathcal{A}λs\lambda_sλcurv\lambda_{\mathrm{curv}}λdiv\lambda_{\mathrm{div}}
A={F1,F2,F3}\mathcal{A}=\{\mathrm{F1},\mathrm{F2},\mathrm{F3}\}1520700
A={DIS∗,DF,F∗}\mathcal{A}=\{\mathrm{DIS}^{*},\mathrm{DF},\mathrm{F}^{*}\}0251200

Table 1 (PDF p. 8).

References

  1. [1]A. Terpin and R. D’Andrea, “Using reinforcement learning to probe the role of feedback in skill acquisition,” arXiv preprint arXiv:2512.08463, 2025.
  2. [2]R. J. Adrian, “Particle-imaging techniques for experimental fluid mechanics,” Annual Review of Fluid Mechanics, vol. 23, no. 1, pp. 261–304, 1991.
  3. [3]M. Raffel, C. E. Willert, F. Scarano, C. J. Kähler, S. T. Wereley, and J. Kompenhans, Particle Image Velocimetry: A Practical Guide. Springer, 2018.
  4. [4]J. Westerweel, G. E. Elsinga, and R. J. Adrian, “Particle Image Velocimetry for Complex and Turbulent Flows,” Annual Review of Fluid Mechanics, vol. 45, no. Volume 45, 2013, pp. 409–436, 2013.
  5. [5]T. Liu and L. Shen, “Fluid flow and optical flow,” Journal of Fluid Mechanics, vol. 614, pp. 253–291, 2008.DOI
  6. [6]T. Corpetti, D. Heitz, G. Arroyo, E. Mémin, and A. Santa-Cruz, “Fluid experimental flow estimation based on an optical-flow scheme,” Experiments in Fluids, vol. 40, no. 1, pp. 80–97, 2006.
  7. [7]Q. Zhong, H. Yang, and Z. Yin, “An optical flow algorithm based on gradient constancy assumption for PIV image processing,” Measurement Science and Technology, vol. 28, no. 5, p. 055208, 2017.DOI
  8. [8]M. G. Santiago, S. T. Wereley, C. D. Meinhar̈t, D. J. Beebe, and R. J. Adrian, “A particle image velocimetry system for microfluidics,” Experiments in Fluids, vol. 25, no. 4, pp. 316–319, 1998.DOI
  9. [9]C. D. Meinhart, S. T. Wereley, and J. G. Santiago, “PIV measurements of a microchannel flow,” Experiments in Fluids, vol. 27, no. 5, pp. 414–419, 1999.
  10. [10]M. Muste, I. Fujita, and A. Hauet, “Large-scale particle image velocimetry for measurements in riverine environments,” Water Resources Research, vol. 44, no. 4, p. W00D19, 2008.DOI
  11. [11]C. J. Legleiter and P. J. Kinzel, “A framework to facilitate development and testing of image-based river velocimetry algorithms,” Earth Surface Processes and Landforms, vol. 49, no. 4, pp. 1361–1382, 2024.
  12. [12]F. Banelli, A. Terpin, A. Bonomi, and R. D’Andrea, “Flow gym: A framework for the development, benchmarking, training, and deployment of flow-field quantification methods,” SoftwareX, vol. 34, p. 102641, 2026.
  13. [13]S. Cai, S. Zhou, C. Xu, and Q. Gao, “Dense motion estimation of particle images via a convolutional neural network,” Experiments in Fluids, vol. 60, no. 4, p. 73, 2019.
  14. [14]T. Kroeger, R. Timofte, D. Dai, and L. Van Gool, “Fast Optical Flow Using Dense Inverse Search,” in Computer Vision – ECCV 2016. Springer International Publishing, 2016, vol. 9908, pp. 471–488.arxiv.org/abs/1603.03590
  15. [15]C. J. Kähler, T. Astarita, P. P. Vlachos, J. Sakakibara, R. Hain, S. Discetti, R. La Foy, and C. Cierpka, “Main results of the 4th International PIV Challenge,” Experiments in Fluids, vol. 57, no. 6, p. 97, 2016.
  16. [16]G. R. Jassal, W. Thielicke, and B. E. Schmidt, “An Optical Flow Algorithm with Automatic Parameter Adjustment for Fluid Velocimetry,” Journal of Open Research Software, vol. 13, p. 22, 2025.DOI
  17. [17]P. Weinzaepfel, J. Revaud, Z. Harchaoui, and C. Schmid, “DeepFlow: Large Displacement Optical Flow with Deep Matching,” in 2013 IEEE International Conference on Computer Vision. IEEE, 2013, pp. 1385–1392.
  18. [18]G. Farnebäck, “Two-Frame Motion Estimation Based on Polynomial Expansion,” in Image Analysis, vol. 2749. Springer Berlin Heidelberg, 2003, pp. 363–370.DOI
  19. [19]S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, p. 1–122, 2011.
  20. [20]J. J. Koenderink, “Theory of Edge-Detection,” in Analysis for Science, Engineering and Beyond, K. Åström, L.-E. Persson, and S. D. Silvestrov, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2012, vol. 6, pp. 35–49.DOI
  21. [21]D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.arxiv.org/abs/1412.6980
  22. [22]Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd ed. Society for Industrial and Applied Mathematics, 2003.DOI
  23. [23]W. L. Briggs, V. E. Henson, and S. F. McCormick, A Multigrid Tutorial, 2nd ed. Society for Industrial and Applied Mathematics, 2000.
  24. [24]M. R. Hestenes and E. Stiefel, “Methods of conjugate gradients for solving linear systems,” Journal of Research of the National Bureau of Standards, vol. 49, no. 6, pp. 409–436, 1952.DOI
  25. [25]P. D. Grontas, A. Terpin, E. C. Balta, R. D’Andrea, and J. Lygeros, “Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers,” in The Fourteenth International Conference on Learning Representations, 2026.
  26. [26]A. Terpin, A. Bonomi, F. Banelli, and R. D’Obb’s, “Synthpix: A lightspeed piv image generator,” SoftwareX, vol. 34, p. 102642, 2026.DOI
  27. [27]G. Bradski, “The OpenCV Library,” Dr. Dobb’s Journal of Software Tools, 2000.
  28. [28]W. Thielicke and E. J. Stamhuis, “PIVlab – Towards User-friendly, Affordable and Accurate Digital Particle Image Velocimetry in MATLAB,” Journal of Open Research Software, 2014.DOI
  29. [29]O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Springer International Publishing, 2015, pp. 234–241.arxiv.org/abs/1505.04597
  30. [30]P. J. Huber, “Robust Estimation of a Location Parameter,” in Breakthroughs in Statistics. Springer New York, 1992, pp. 492–518.
  31. [31]N. Parikh and S. Boyd, “Proximal algorithms,” Found. Trends Optim., vol. 1, no. 3, p. 127–239, 2014.DOI
  32. [32]S. Baker and I. Matthews, “Lucas-Kanade 20 Years On: A Unifying Framework,” International Journal of Computer Vision, vol. 56, no. 3, pp. 221–255, 2004.
  33. [33]——, “Equivalence and efficiency of image alignment algorithms,” in Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, vol. 1, 2001, pp. I–I.

Paper details

vlnc:W_4a6cgpvr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents