Efficient Event-based MeshFlow (EEMFlow)

• The paper introduces Efficient Event-based MeshFlow (EEMFlow) as a novel method for predicting spatially smooth, sparse motion fields from event cameras.
• It uses a lightweight encoder-decoder network with Confidence-induced Detail Completion (CDC) and Adaptive Density Module (ADM) to preserve sharp motion boundaries and optimize event density.
• Quantitative results on the HREM(+) dataset demonstrate low end-point errors (EPE 5.05) and high efficiency (142.9 FPS) with a compact 1.24M parameter model.

Efficient Event-based MeshFlow (EEMFlow) is a method for event-based meshflow estimation, a task defined as predicting a spatially smooth sparse motion field from event cameras. It was introduced together with the High-Resolution Event Meshflow (HREM) dataset and later extended to HREM+, which varies event density. The framework comprises a lightweight encoder-decoder network for meshflow, an upgraded variant for dense event optical flow, a Confidence-induced Detail Completion (CDC) module to preserve sharp motion boundaries, and an Adaptive Density Module (ADM) to adjust input event density to a more optimal range. The stated motivation is twofold: the lack of meshflow-specific event datasets and methods, and the underexplored challenge of event data density (Luo et al., 5 Oct 2025).

1. Conceptual scope and problem formulation

EEMFlow targets event-based meshflow estimation rather than dense optical flow alone. In the formulation used by the method, meshflow is a spatially smooth, sparse motion field defined at mesh (grid) vertices. This differs from dense optical flow, which is defined over the image lattice. The distinction is operational in EEMFlow: the base network directly regresses the full meshflow at mesh resolution, whereas the upgraded EEMFlow+ variant supports dense event optical flow through a coarse-to-fine residual refinement pipeline (Luo et al., 5 Oct 2025).

The work frames meshflow as a representation suited to event cameras, which provide temporally precise asynchronous measurements but also exhibit substantial variation in event density. The paper explicitly identifies two research gaps: the lack of meshflow-specific event datasets and methods, and the underexplored challenge of event data density. A common misconception is that EEMFlow is only an optical-flow network with a different decoder. The description of the task and the output structure indicates otherwise: meshflow is treated as its own estimation target, with a dedicated dataset, labels, and evaluation protocol (Luo et al., 5 Oct 2025).

A plausible implication is that the method is intended for downstream motion-aware alignment tasks where a spatially smooth field is preferable to a dense but potentially noisier displacement map. That implication is consistent with the reported use of EEMFlow results for image alignment with reduced ghosting.

2. HREM and HREM+: dataset design and supervision

The High-Resolution Event Meshflow (HREM) dataset is described as the first meshflow-focused event dataset and as the highest among available event datasets in resolution, at 1280 × 720. Its scale is specified as 100 virtual scenes (indoor/outdoor), with 20,000 training and 8,000 test samples. The scenes simulate complex motions and dynamic objects using Blender, and each sample provides both meshflow and dense optical flow annotations (Luo et al., 5 Oct 2025).

The event streams are generated with ESIM, V2E, DVS-Voltmeter, which the paper characterizes as advanced event data simulators. The extension HREM+ modulates event generation thresholds to produce varying densities across a uniform density range [0.05, 0.95]. This is not an ancillary addition: the density-controlled extension is central to the paper’s study of robustness under varying event densities (Luo et al., 5 Oct 2025).

The meshflow ground truth is produced by downsampling dense optical flow. The stated procedure is that dense optical flow is downsampled by propagating each cell’s center motion to grid vertices, after which median filters are used in two stages to suppress noise and yield a spatially smooth field. This directly encodes the structural prior behind meshflow: smoothness is enforced at the supervisory level, not only in the decoder or loss (Luo et al., 5 Oct 2025).

The dataset contributions are significant because earlier efficient event-based flow methods emphasized either temporal continuity, direct event-trace processing, or normal-flow intermediates, but not a dedicated high-resolution meshflow benchmark. TMA focused on temporally fine-grained motion information for event optical flow (Liu et al., 2023), while IDNet removed explicit correlation volumes for efficient dense flow (Wu et al., 2022). EEMFlow’s dataset work therefore addresses a different gap: standardized supervision for meshflow-specific learning (Luo et al., 5 Oct 2025).

3. Network architecture and meshflow estimation pipeline

EEMFlow is specified as an encoder-decoder model tailored for meshflow. Its input consists of two consecutive event voxel grids, exemplified as $V_{t_{k-1}\to t_k}$ and $V_{t_k\to t_{k+1}}$, representing binned event data. The encoder is a three-level pyramid with shared-weight convolutions for multi-scale feature extraction. Motion correspondence is handled by Dilated Feature Correlation (DFC), which is designed to capture a large receptive field but low computational cost. The decoder uses group shuffle convolutions and directly regresses the meshflow at mesh resolution (Luo et al., 5 Oct 2025).

The DFC design is central to the efficiency claim. The search grid is described as being sampled densely near the center and sparsely at the periphery, which reduces operation count while supporting large search radii. The decoder choice is likewise explicitly motivated by efficiency: group shuffle convolutions, inspired by ShuffleNet, are used to minimize model size (Luo et al., 5 Oct 2025).

For dense optical flow, the network is upgraded to EEMFlow+, which employs coarse-to-fine residual refinement. Intermediate flow is upsampled, and corrections are added at each scale. In this sense, EEMFlow and EEMFlow+ separate the sparse-smooth mesh representation from dense flow recovery rather than conflating them in a single output head (Luo et al., 5 Oct 2025).

This architectural choice places EEMFlow in a broader efficiency-oriented lineage, but with a distinct design point. TMA retains a RAFT-style backbone and improves optical flow by exploiting temporal continuity across event splits (Liu et al., 2023). IDNet instead obviates correlation volumes and estimates flow directly from event traces with a ConvGRU backbone (Wu et al., 2022). STSSM replaces iterative refinement and cost volumes with a Spatio-Temporal State Space Model (Humais et al., 9 Jun 2025). EEMFlow differs by optimizing a meshflow-centric encoder-decoder with DFC and shuffle-based decoding, while also explicitly addressing density variation (Luo et al., 5 Oct 2025).

4. CDC and ADM: dense completion and density adaptation

The Confidence-induced Detail Completion (CDC) module is introduced for dense optical flow to address blurred edges in upsampled optical flow. It contains a self-corrector branch, in which dense convolutional layers refine flow locally, and a self-correlation branch, in which self-attention identifies distant, similar-motion features to correct errors. The module learns a confidence map $W^i$ that combines bilinear upsampling and corrections end-to-end (Luo et al., 5 Oct 2025).

The completion equations are stated as

$$\tilde{F}^i= \alpha \,\mathcal{W}(\bar{F}^i, \Delta F^i) + (1-\alpha) (A^i \otimes \bar{F}^i)$$

and

$$F^i_\uparrow= W^i \odot \bar{F}^i + (1-W^i) \odot \tilde{F}^i,$$

where $\bar{F}^i$ is the bilinearly upsampled flow, $\Delta F^i$ the correction, $A^i$ attention weights, and $\alpha$ a fixed mixing coefficient empirically set, for example, to 0.6 (Luo et al., 5 Oct 2025).

The Adaptive Density Module (ADM) is introduced because event flow estimation performance depends strongly on input event density, and because different cameras and settings yield widely varied densities. ADM is placed before the main network and trained jointly with it. It contains a Multi-Density Changer (MDC), an encoder-decoder with 3 levels that produces multi-scale density-adjusted representations, and a Multi-Density Selector (MDS) that generates pixelwise (softmax-weighted) selection weights to fuse the MDC output and the input event representation adaptively (Luo et al., 5 Oct 2025).

The losses are described as a multi-scale L1 loss (Eq. 16) and a density-matching loss (Eq. 17). The stated purpose is to adjust the density of input event data to a range optimal for downstream flow/meshflow networks. The density ablations report an optimal event density range for training ([0.45, 0.55]), and the claim is that ADM adaptively brings arbitrary input densities to this zone (Luo et al., 5 Oct 2025).

A common misunderstanding is that ADM is merely a data-augmentation strategy. In the formulation given here, ADM is a learnable plug-in module integrated into the estimation pipeline and optimized jointly with the downstream network, not a fixed preprocessing heuristic.

5. Quantitative results, efficiency, and qualitative behavior

On the HREM(+) Dataset, EEMFlow is compared to EV-FlowNet, ERAFT, GMA, DPFlow, etc. The reported result is that EEMFlow achieves the lowest average EPE — 5.05 (with ADM, dt=1). Its runtime reaches 142.9 FPS, it is stated to be 30× faster than modern optical-flow baselines at comparable or better accuracy, and the model size is 1.24M params (Luo et al., 5 Oct 2025).

For dense optical flow on the DSEC benchmark, EEMFlow+ achieves $\text{1PE}=11.4$ and $V_{t_k\to t_{k+1}}$0, with a significant speed advantage against all high-performing event-based flow models. The paper further states that ADM helps to significantly improve the performance of EEMFlow and EEMFlow+ by 8% and 10%, respectively, and that on HREM+ it improves optical flow by up to 14% for different models (Luo et al., 5 Oct 2025).

The ablation results attribute the efficiency/accuracy tradeoff to specific components: DFC and channel shuffle convolution are shown to offer the best tradeoff between efficiency and accuracy; CDC restores sharp, accurate edge motion, and removing either CDC branch reduces accuracy; the full CDC yields the best results with only modest computational cost (Luo et al., 5 Oct 2025).

The reported qualitative behavior is also task-specific. EEMFlow produces smoother, globally consistent meshflow fields, and resulting image alignments with EEMFlow meshflow have reduced “ghosting” (color fringes), which the source identifies as critical for applications such as high-dynamic range imaging or video stabilization (Luo et al., 5 Oct 2025).

Setting	Reported result	Source detail
HREM(+), meshflow	EPE 5.05	with ADM, dt=1
HREM(+), runtime	142.9 FPS	30× faster
Model size	1.24M params	smallest model size
DSEC, EEMFlow+	1PE = 11.4	FPS = 39.2
ADM gain	8% / 10%	EEMFlow / EEMFlow+

6. Robustness, generalization, and relation to adjacent event-based motion methods

The robustness analysis emphasizes density variation. HREM+ is explicitly designed for a thorough study of the robustness of existing methods across data with varying densities, and the reported conclusion is that performance depends strongly on input density. The paper states that pretraining on HREM+ and using ADM enables SOTA generalization to real data (MVSEC), outperforming baselines trained solely on real data (Luo et al., 5 Oct 2025).

This density-centric framing differentiates EEMFlow from other efficiency-oriented event-motion methods. TMA focuses on temporal continuity by splitting event streams into multiple fine-grained segments, aligning them with a linear lookup strategy, and aggregating them with a motion pattern aggregation module; it reports a 6% improvement in accuracy and a 40% reduction in inference time on DSEC-Flow relative to E-RAFT (Liu et al., 2023). IDNet instead argues that event traces provide a natural search direction, thereby avoiding explicit correlation volumes; its base model uses 80% fewer parameters, 20x less memory footprint, and runs 40% faster on NVIDIA Jetson Xavier NX, while its online TID mode offers 8 ms ultra-low latency (Wu et al., 2022). STSSM replaces cost volumes and iterative refinement with a state-space model and achieves 4.5x faster inference and 8x lower computations compared to TMA (Humais et al., 9 Jun 2025).

A broader methodological contrast also appears in the literature on normal flow. “Motion and Structure from Event-based Normal Flow” formulates geometric estimation directly from the constraint $V_{t_k\to t_{k+1}}$1, emphasizing linear and continuous-time solvers rather than learned dense correspondences (Ren et al., 2024). “Real-time Motion Segmentation with Event-based Normal Flow” uses dense normal flow as an intermediate representation for graph-cut segmentation and reports nearly a 800x speedup over an open-source baseline (Zhong et al., 24 Feb 2026). Earlier event-driven approaches such as ARMS addressed the aperture problem with multi-scale plane fitting and supported event-by-event motion estimation and predictions of up to 500 ms (Akolkar et al., 2018). Hardware-oriented work such as EventShiftFlow trades dense sub-pixel flow for sparse, quantized velocity estimates, requiring less than 2 kB of storage and using no DSP blocks (Bizzi et al., 27 May 2026).

Within this landscape, EEMFlow is best understood as a meshflow-centered synthesis of three concerns: a dedicated supervisory substrate (HREM/HREM+), a lightweight architecture for sparse smooth motion estimation, and an explicit mechanism for density adaptation. This suggests that its principal contribution is not only efficiency in the narrow runtime sense, but also the formalization of event-based meshflow as a dataset-backed, density-aware estimation problem (Luo et al., 5 Oct 2025).