---
title: High-Resolution Event Meshflow (HREM)
url: https://www.emergentmind.com/topics/high-resolution-event-meshflow-hrem
type: topic
---

# High-Resolution Event Meshflow (HREM)

Searching arXiv for the cited HREM-related papers to ground the article.
High-Resolution Event Meshflow (HREM) denotes a recent family of formulations for recovering spatially coherent, higher-resolution motion fields from event-driven or otherwise coarse flow observations. In its most explicit usage, HREM is a high-resolution event-camera dataset and supervision pipeline for event-based meshflow and optical flow at \(1280\times720\), accompanied by the Efficient Event-based MeshFlow (EEMFlow) network, the Confidence-induced Detail Completion (CDC) module, and the HREM+ multi-density extension [2510.04111]. In the supplied literature, the same acronym is also used for a mesh-based super-resolution architecture derived from multiscale graph neural networks [2409.07769] and for a reduced-order reconstruction framework for real-time Event-Based Imaging Velocimetry (rt-EBIV) [2605.04186]. The shared technical theme is super-resolution or refinement of flow information under sparsity, reduced sampling, or event-based sensing.

## 1. Terminology and scope

The designation **HREM** is not used in a single, fully standardized way across the supplied literature. Instead, it appears in three closely related senses: as a dataset for event-based meshflow learning, as a graph-based super-resolution architecture on mesh elements, and as a reduced-order real-time reconstruction framework for event-based velocimetry. This suggests that the acronym functions more as a thematic label for high-resolution flow recovery than as a unique canonical method.

| Usage of HREM | Paper | Core object |
|---|---|---|
| High-Resolution Event Meshflow dataset and learning pipeline | [2510.04111] | Dataset, EEMFlow, CDC, HREM+ |
| HREM super-resolution architecture | [2409.07769] | Synchronized message passing, CSP, unpooling, FSP |
| HREM framework for rt-EBIV | [2605.04186] | POD mapping, linear dynamics, KF/LSE/LSE+VR |

In the event-camera setting, meshflow is defined as a **sparse, spatially smooth motion field defined at the vertices of a regular grid**. Compared with dense optical flow, it is presented as a representation that **dramatically reduces redundancy and computational cost while still capturing non-rigid and multi-planar motion** [2510.04111]. In the fluid and velocimetry settings, HREM denotes super-resolution of flow fields from coarse or reduced observations rather than a particular event-camera dataset [2409.07769].

## 2. Dataset construction and meshflow formulation

In [2510.04111], HREM was created to enable supervised learning of event-based meshflow and optical flow at high resolution, in the presence of **dynamic objects, complex motions, and varying illumination**. The dataset statistics reported are **100 virtual scenes rendered in Blender (40 indoor, 60 outdoor)**, **20,000 training pairs**, **8,000 test pairs**, and **spatial resolution: \(1280\times720\)**. Four test sub-sequences are defined by scene type and motion speed: **Outdoor_slow (mean motion 0–30 px)**, **Outdoor_fast (30–100 px)**, **Indoor_slow (0–20 px)**, and **Indoor_fast (20–100 px)**. Event data are captured at **two temporal windows: \(dt=1\) (60 Hz meshflow) and \(dt=4\) (15 Hz)** [2510.04111].

The synthesis pipeline begins with **camera trajectories \(\Gamma(t)\) with 6-DoF generated via PyBullet**. Blender renders **high-frame-rate videos \(I(u,\tau_k)\)** and **ground-truth dense optical flow \(F_{\tau_k\rightarrow\tau_{k+1}}\)**. Meshflow ground truth is then derived by placing **a \(16\times16\) regular mesh on each image**, collecting a dense flow vector at each cell center, propagating that vector to the **\(3\times3\) block of nearby vertices**, and applying two median filters: a first median filter \(f_1\) on each vertex and a second \(3\times3\) median filter \(f_2\) on the resulting sparse field to enforce spatial smoothness. The output is **a sparse \(16\times16\) motion field \(MF\)**, bilinearly upsampled for display [2510.04111].

Event streams are synthesized by **three event simulators (ESIM, V2E, DVS-Voltmeter)**. Each event \(e=(u_e,t_e,p_e)\) satisfies
$$
L(u_e,t_e)-L(u_e,t_e-\Delta t)\ge p_e C,
$$
where \(L=\ln I\), \(p_e\in\{\pm1\}\), and \(C\) is the contrast threshold. To select among simulators, the pipeline computes the **Image of Warped Events (IWE) contrast**,
$$
IWE(u,t_{ref})=\sum_e\delta(u-u'_e),\qquad
u'_e=u_e + \frac{t_{ref}-t_e}{t_j-t_i}\cdot F_{i\rightarrow j}(u_e),
$$
and chooses the simulator with maximum \(\mathrm{Var}(IWE)\) [2510.04111].

A further ingredient is event density. HREM+ extends HREM by synthetically varying the event-generation threshold \(C\) to produce event data densities \(D\in[0.05,0.95]\). Density is defined from a voxel grid \(V(b,u)\) over \(B\) temporal bins as
$$
D=\frac{1}{HW}\sum_u 1\left\{\sum_b |V(b,u)|>0\right\}.
$$
This explicit parameterization of density is central to the later ADM module and to the paper’s robustness study across varying event densities [2510.04111].

The meshflow representation itself is formulated on an \(M\times N\) mesh with motion vectors \(M_{i,j}\in\mathbb{R}^2\) at each vertex. The stated goal is a sparse field that is both **consistent with the underlying dense flow \(F\)** and **spatially smooth**. For clarity, the two median filters may be viewed as minimizing
$$
L_{smooth}=\sum_{(i,j)}\sum_{(i',j')\in N(i,j)} \|M_{i,j}-M_{i',j'}\|_1,
$$
subject to matching propagated candidate vectors, although the implementation uses the two cascaded \(3\times3\) median filters rather than explicit optimization [2510.04111].

## 3. EEMFlow architecture for event-based meshflow

The principal learning architecture associated with HREM is **Efficient Event-based MeshFlow (EEMFlow)**. Its inputs are **two voxelized event volumes \(V_{k-1\rightarrow k},V_{k\rightarrow k+1}\in\mathbb{R}^{B\times H\times W}\)**. The network has three major stages: a pyramid encoder, Dilated Feature Correlation (DFC), and feature fusion with a lightweight decoder [2510.04111].

The **Pyramid Encoder** has **three levels \(i=0,1,2\)** operating at downsample factors **\(1/4,1/8,1/16\)**. Each level uses **shared \(3\times3\) convolutions** to extract features \(F^i_{k-1},F^i_k\). At each scale, **Dilated Feature Correlation** forms a cost volume
$$
C^i(u,d)=\langle F^i_k(u),F^i_{k-1}(u+d)\rangle,\qquad d\in\mathcal{B},
$$
where \(\mathcal{B}\) is a sparse “dilated” search grid,
$$
\mathcal{B}=\{(d_x,d_y):|d_x|+|d_y|\ne 2k \text{ for all } 2\le k\le r\}.
$$
This masking preserves a large receptive field while lowering cost. The reported comparison is that standard correlation with radius \(r\) has cost \(O((2r+1)^2HW)\), whereas DFC achieves the same range with **approximately 50% fewer operations** [2510.04111].

The decoder stacks the cost volumes with pooled features and uses grouped **“Shuffle” convolutions** to regress a **full-resolution meshflow \(MF_{k\rightarrow k+1}\)**. The paper’s ablations state that **Shuffle conv vs. standard/depthwise** yields the **best trade-off in decoder**, and that **DFC vs. standard correlation** provides **large radius at low cost** [2510.04111].

This architecture is explicitly lightweight. The reported meshflow model has **1.24 M parameters**, **7 ms/inference (142.9 FPS)**, and **Avg EPE = 5.50 (\(dt=1\)), 24.45 (\(dt=4\))** on HREM. Relative to **ERAFT [Gehrig et al. 2021]**, the paper reports **\(-76\%\) params, \(-92\%\) time, \(-8.3\%\) EPE** [2510.04111].

## 4. Dense-flow extension, CDC, and density adaptation

When EEMFlow is upgraded to predict dense optical flow, the architecture uses a **coarse-to-fine residual refinement** and introduces the **Confidence-induced Detail Completion (CDC)** module at each upsampling stage [2510.04111]. At decoder level \(i\), after bilinear upsampling of the current flow estimate \(F^i\), the network warps features, forms a correlation, and regresses a residual:
$$
F^{i+1}=Conv^i(C(F^i_{\uparrow},\ldots))+F^i_{\uparrow}.
$$

CDC takes the upsampled flow \(\bar F^i=\mathrm{Upsample}(F^i)\) and processes it through two branches. The **Self-Corrector branch** is a **5-layer dense conv net** on concatenated features \([F^i_{k-1},F^i_k]\) that predicts a residual \(\Delta F^i\) and a confidence map \(W^i\in[0,1]\). The **Self-Correlation branch** is **a small self-attention module** that computes weights \(A^i\) to detect feature-space patterns matching blurring errors in \(\bar F^i\). The corrections are combined as
$$
\tilde F^i=\alpha\cdot \mathrm{Warp}(\bar F^i,\Delta F^i) + (1-\alpha)(A^i\odot \bar F^i),
$$
with **\(\alpha=0.6\) empirically**, and the final upsampled flow is
$$
F^i_{\uparrow}=W^i\odot \bar F^i + (1-W^i)\odot \tilde F^i.
$$
The ablations report improvements from **self-corrector alone (0.81\(\rightarrow\)0.79 EPE)**, **self-correlation alone (0.89\(\rightarrow\)0.79)**, and **both (0.79\(\rightarrow\)0.75)** at **39.2 FPS** [2510.04111].

The second major extension is **HREM+**, which contains events at **uniformly sampled densities \(D\in[0.05,0.95]\)**. The stated empirical observation is that **models trained at density \(\approx0.5\) perform best across real datasets**. To generalize across densities, the paper introduces the **Adaptive Density Module (ADM)**. Its **Multi-Density Changer (MDC)** is **a 3-level encoder–decoder** that maps an input \(V\) to \(\{V^{MDC}_1,V^{MDC}_2,V^{MDC}_3\}\), while the **Multi-Density Selector (MDS)** concatenates the original \(V\) and \(V^{MDC}\), applies **a two-layer conv**, and produces pixel-wise softmax weights \(\{w,1-w\}\). The adjusted volume is
$$
V^{ad}=w\cdot V + (1-w)\cdot V^{MDC},
$$
which is then fed to EEMFlow or EEMFlow+ [2510.04111].

Training uses **\(L_1\) regression at full resolution** for both meshflow and optical flow:
$$
L_{flow}=\sum_u \|F_{pred}(u)-F_{GT}(u)\|_1.
$$
For ADM, two auxiliary losses on HREM+ are defined using moderate-density \(V^{GT}\):
$$
L_{MDC}=\sum_{k=1}^3 \sqrt{(V^{MDC}_k-V^{GT}_k)^2+\xi^2},\qquad \xi=10^{-3},
$$
and
$$
L_{MDS}=\|D(V^{ad})-D(V^{GT})\|_1.
$$
The total loss is
$$
L_{total}=L_{flow}+\lambda_1 L_{MDC}+\lambda_2 L_{MDS},
$$
with **\(\lambda_1=0.1\)** and **\(\lambda_2=10\)** [2510.04111].

## 5. Reported empirical performance and ablations

The principal reported results for HREM concern meshflow estimation on the HREM dataset and optical-flow estimation on DSEC. For meshflow on HREM, **EEMFlow** achieves **Avg EPE = 5.50 (\(dt=1\))** and **24.45 (\(dt=4\))**, while **with ADM** these improve to **5.05 (\(dt=1\), \(-8.2\%\))** and **22.43 (\(dt=4\), \(-8.3\%\))** [2510.04111].

| Setting | Metric summary | Reported runtime |
|---|---|---|
| EEMFlow on HREM | Avg EPE \(=5.50\) (\(dt=1\)), \(24.45\) (\(dt=4\)) | 7 ms, 142.9 FPS |
| EEMFlow + ADM on HREM | Avg EPE \(=5.05\) (\(dt=1\)), \(22.43\) (\(dt=4\)) | not separately stated |
| EEMFlow+ + both CDC branches on DSEC | 1PE \(=11.4\), EPE \(=0.75\), AE \(=2.67\) | 39.2 FPS |
| EEMFlow+ + ADM on DSEC | 1PE \(=10.3\), 2PE \(=3.68\), 3PE \(=2.12\), EPE \(=0.73\), AE \(=2.70\) | 22.9 FPS |

For dense optical flow on DSEC, **EEMFlow+ without CDC** is reported at **60.4 FPS**, **1PE = 15.5**, **EPE = 0.89**, and **AE = 3.11**. Adding **both CDC branches** yields **39.2 FPS**, **1PE = 11.4**, **EPE = 0.75**, and **AE = 2.67**, described as a **state-of-the-art speed–accuracy trade-off**. With **ADM**, the reported metrics become **22.9 FPS**, **1PE = 10.3 (\(-9.7\%\))**, **2PE = 3.68 (\(-6.3\%\))**, **3PE = 2.12 (\(-1.5\%\))**, **EPE = 0.73**, and **AE = 2.70** [2510.04111].

The density-generalization results are central to the HREM+ extension. The reported finding is that **training on HREM+ densities \(D\in[0.45,0.55]\) achieves the best cross-dataset performance on MVSEC and DSEC**. In addition, **spatial/temporal guided subsampling on DSEC confirms that lower \(D\) degrades performance linearly; ADM mitigates up to 60% density reduction**. The paper further reports that **ADM also yields +8–14% gains when plugged into TMA, BFlow, etc.** [2510.04111].

Among ablations, the paper states that **softmax fusion in ADM outperforms average/direct outputs**, and that **bilinear upsampling of meshflow yields better global alignment than mesh-homography or kriging**. In the conclusions, HREM+ is described as **the first high-resolution (\(1280\times720\)), dynamic, multi-density event dataset with both meshflow and dense optical flow ground truth**, while EEMFlow is reported as achieving **SOTA accuracy (Avg EPE = 5.50) at 142.9 FPS**. The abstract additionally characterizes EEMFlow as **30x faster** than a recent state-of-the-art flow method and states that ADM improves the performance of **EEMFlow and EEMFlow+ by 8% and 10%, respectively** [2510.04111].

The applications named in the paper are **real-time video stabilization, SLAM, HDR imaging in challenging lighting, and autonomous navigation**. The listed future directions are **end-to-end fusion of images and events**, **learning to synthesize optimal densities on the fly**, and **extending meshflow to 3D scene flow with event arrays** [2510.04111].

## 6. Related HREM formulations in fluid super-resolution and rt-EBIV

In the supplied literature, HREM also labels a **four-stage super-resolution architecture** built from **synchronized message passing**, a **coarse-scale processor (CSP)**, **graph-unpooling**, and a **fine-scale processor (FSP)**, summarized from Barwey et al.’s work on mesh-based fluid-flow super-resolution [2409.07769]. In that formulation, synchronized message passing modifies a baseline graph-convolution layer by averaging aggregated features across coincident nodes:
$$
a_j^{\,l}=\frac{1}{|\mathcal{C}(j)|}\sum_{i\in\mathcal{C}(j)} a_i^{\,l},
$$
with node and edge updates performed by \(\mathrm{MLP}_y^l\) and \(\mathrm{MLP}_e^l\). The two MLPs are reported to employ **ELU activations and layer-norm, and residual connections inside the MLP itself**. The **CSP** embeds a query coarse element graph and up to **26 neighbors in a structured hexahedral stencil** into a latent representation using **\(C\) synchronized message-passing layers**. A **KNN-based interpolation** with **\(K=8\)** performs graph-unpooling from coarse to fine latent graphs, and the **FSP** applies **\(F\) standard (unsynchronized) message-passing layers** as a **fine-scale corrector** before decoding a residual velocity field in \(\mathbb{R}^3\) [2409.07769].

The numerical experiments reported in that summary are restricted to **Taylor–Green Vortex in a periodic cube**, solved by **NekRS with spectral-element hexahedral mesh**, at **Re = 1600 and 3200**. Training uses **three time-levels near peak dissipation (\(t=8,9,10\))**, coarsened from **\(P=7\) DNS to \(P=1\)**. Two model variants are given: **Model 1 (pure coarse-scale): \(C=12\), \(F=0\)** and **Model 2 (multiscale): \(C=6\) coarse layers + \(F=6\) fine layers**, with neighborhood sizes **0, 6 or 26 coarse-element neighbors**. The reported findings are that **both models dramatically outperform classical spectral-element interpolation**, that at **Re = 1600** the multiscale variant with neighbors reduces **element-MSE by approximately 20–30%**, that at **Re = 3200** the purely coarse-scale model is **nearly as accurate**, and that **errors scale roughly with Reynolds number and concentrate where coarse field has high local variance**. The same summary states that **time-extrapolation** and **Reynolds-extrapolation** remain reasonable, though **upward Re-extrapolation is more challenging** [2409.07769].

A distinct HREM formulation appears in **“Real-Time Estimation of High-Resolution Flow Fields and Reduced-Order Coordinates from Event-Based Imaging Velocimetry”** [2605.04186]. There, HREM couples a **very fast, low-resolution event-based velocimetry pipeline** with an **offline-trained mapping and dynamical model** that reconstructs high-resolution velocity fields in real time. Offline, paired LR and HR fields are used for **Proper Orthogonal Decomposition (POD)** on both resolutions, yielding \(\Phi_{LR}\) and \(\Phi_{HR}\), and for identification of an **LR\(\rightarrow\)HR linear mapping**
$$
M=A_{HR}A_{LR}^T\left(A_{LR}A_{LR}^T+\lambda I_n\right)^{-1},
$$
together with a **one-step linear dynamical operator**
$$
F=A_{HR}^{+}\left(A_{HR}^{-}\right)^\dagger.
$$
Online, each LR snapshot is projected onto the LR basis, an estimated HR coefficient vector is produced, and the HR field is reconstructed by
$$
\hat{\mathbf{u}}_{HR}(\mathbf{x},t)=\Phi_{HR}\Sigma_{HR}\hat a_{HR}(t)+\overline{\mathbf{U}}(\mathbf{x}).
$$

That rt-EBIV HREM compares **three estimators**: a **direct Kalman filter (KF)**, **Linear Stochastic Estimation + KF (LSE)**, and **Variance-Rescaled LSE + KF (LSE+VR)**. The computational costs per snapshot are stated as **\(\mathcal{O}(rN_p^{LR})\)** for LR projection, **\(\mathcal{O}(r^2)\)** for reduced-state estimation, and **\(\mathcal{O}(rN_p^{HR})\)** for HR field reconstruction, with the reconstruction step dominating. On the jet dataset at **\(r\approx200\)**, a **C++ microbenchmark** reports that the added time per frame remains **below 0.8 ms in the worst case**, and that if only reduced coordinates are needed, the latent pipeline runs in **approximately \(10\,\mu s\)**, described as negligible compared with raw cross-correlation latency. Quantitatively, on the jet case, normalized RMSE is reported as **cubic interpolation \(\approx0.088\)**, **KF \(0.079\)**, **LSE \(0.069\)**, and **LSE+VR \(0.070\)**; on the channel flow, **cubic \(\sim0.12\)** versus **LSE \(\sim0.104\)**. The paper also reports that HREM better preserves **turbulent kinetic energy**, **temporal spectra**, **spatial spectra**, and **temporal autocorrelation** than cubic interpolation [2605.04186].

A common misconception would be to treat these three HREM instances as interchangeable. The supplied literature does not support that interpretation. Rather, HREM refers to a set of related high-resolution flow-recovery strategies: one centered on event-camera meshflow datasets and lightweight neural estimation [2510.04111], one on graph-based mesh super-resolution in computational fluid dynamics [2409.07769], and one on reduced-order real-time reconstruction for event-based velocimetry [2605.04186]. What unifies them is the objective of reconstructing dynamically meaningful, higher-resolution flow structure from sparse, localized, or low-resolution inputs.

Source: https://www.emergentmind.com/topics/high-resolution-event-meshflow-hrem