WorksComputer Science
Event-Based Optical Flow Constraint via Time-Sliced Plane Fitting
Abstract
Event-based optical flow estimation using spatiotemporal planar approximation inherently assumes constant velocity and tends to average out temporal velocity variations. In this paper, we propose a flow estimation method that extends this planar approximation along the temporal axis. The resulting event-based optical flow constraint allows existing variational optical flow formulations to model temporal motion variation. Specifically, we partition the space-time volume into sub-intervals and apply a two-step planar approximation to track object edges consistently. Moreover, a spatial compensation mechanism and a temporal regularization term are incorporated to compensate for edge displacements and temporally stabilize the flow estimation, respectively. The experimental results show that modeling temporal motion variation improves flow estimation accuracy over conventional methods.
Introduction
An event camera is a bio-inspired vision sensor that asynchronously detects per-pixel brightness changes, outputting events with timestamp , pixel coordinates , and the polarity of the brightness change . Due to their high temporal resolution on the order of s, high dynamic range (HDR) and low power consumption [1], they are increasingly utilized in applications such as object detection [2], HDR video reconstruction [3], and edge systems [4, 5].
Optical flow estimation determines apparent displacements between time points, and is applied in fields such as video compression [6] and object detection [7]. As event cameras fundamentally differ from conventional frames, existing frame-based approaches are not directly applicable. To estimate flow without frame conversion, methods such as contrast maximization [8] and spatiotemporal planar approximation of the event point cloud [9] have been proposed.
A previous study [9] estimates optical flow via planar approximation of the event point cloud, but this formulation inherently enforces a velocity constancy assumption within the temporal estimation window, yielding only an average velocity field. This limitation may be particularly problematic in practical scenarios. Even when objects move at constant velocity in 3D space, their projected motion on the image plane can exhibit apparent acceleration—for instance, during camera translation, points near the Focus of Expansion (FOE) move slowly while peripheral points move faster. Such spatially varying apparent acceleration cannot be captured under a constant-velocity assumption, motivating the need for an acceleration-aware estimation method.
In this paper, we partition the space-time volume in the temporal direction and estimate flow for each sub-interval, enabling the model to account for apparent acceleration during flow estimation arising from both object motion and perspective projection.
This paper is an extended version of a non-peer-reviewed technical report [10].
Related Work
Event-based optical flow estimation methods can be broadly categorized into learning-based and model-based approaches [1]. Learning-based methods (e.g., EV-FlowNet [11], E-RAFT [12]) achieve high accuracy but require extensive training and lack geometric interpretability. Among model-based methods, the contrast maximization framework [8] is widely used, which estimates motion parameters by maximizing the sharpness of motion-compensated event images.
Alternatively, another class of model-based approaches [9] explicitly leverages the local geometric structure of events in space-time. By extracting spatiotemporal gradients directly from event surfaces, this approach uniquely allows the integration of classical, mathematically well-founded optical flow frameworks (such as the Horn–Schunck method [13]) without requiring pseudo-frame generation. Specifically, a previous study [9] assumes that when an object edge moves, the resulting event point cloud forms a surface in space-time along its trajectory. Optical flow is then estimated by approximating this surface as a single plane via RANSAC (Fig. 1), applied to each spatially partitioned sub-region. However, this formulation shares the constant-velocity assumption within the temporal estimation window. The proposed method addresses this limitation by extending the planar approximation along the temporal axis.
This work was supported by JSPS KAKENHI Grant Number JP24K14996. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Accepted for publication at IEEE VCIP 2026.

Figure 1. Concept of optical flow estimation based on spatiotemporal planar approximation. In subsequent figures, the distinction between positive and negative polarities is omitted, and only one polarity is illustrated.
Preliminaries
For each space-time sub-region with , where is the image domain, approximating the event point cloud with a plane yields two intersection lines, at and at . These two lines are related by translation, and their motion can be represented on the solution space of and as shown in the left panel of Figure 2. The displacement between these lines divided by the time interval corresponds to the target flow velocity for . The relationship between the coefficients of the plane equation and the flow components is given by
Figure 2. Relationship between edge-line translation and optical flow constraint for at : (left) constraint line from edge-line translation, (right) constraint line from brightness constancy equation.
Here, within each region , the functions , and are constant on the line and zero elsewhere.
Similarly, for a frame image , the brightness constancy constraint equation based on the assumption of constant intensity over an infinitesimal time interval is expressed as
From this, the solution for the flow from the frame image lies on the line shown in the right panel of Figure 2.
Therefore, from the correspondence between Eqs. (1) and (2), the coefficient functions of the plane equation formed by the events in each sub-region correspond to the spatiotemporal gradients of the frame image brightness, respectively. Although the coefficients can be obtained from each polarity, the two polarities can be considered as a two-channel image. Then, following the Horn–Schunck type method [13], the optical flow is estimated by minimizing an objective functional combining a data term (2) and a spatial smoothness term, defined as
where denotes color channels (i.e., two polarities), is a positive weight between channels, and is a positive value corresponding to the weight between terms [14]. The flow minimizing (3) is defined as the estimated flow.
Proposed Method
To capture temporal velocity variations, we extend the conventional method [9] along the temporal axis and propose a flow estimation method that accounts for acceleration.
Temporal Division and Two-Step Planar Approximation
As illustrated in Figure 3, the proposed method divides the total estimation interval into sub-intervals (). For each sub-interval, the enclosed events are extracted, and spatiotemporal planar approximation is performed to estimate the local optical flow, thereby capturing temporal variations of the event surfaces. The resulting velocity field is then associated with the central timestamp of each sub-interval.

Figure 3. Overview of the proposed temporal division and sub-interval planar approximation.
If planar approximation is performed independently within each short sub-interval, there is a risk of detecting different planes (motions) between consecutive intervals due to noise or interference from adjacent surfaces. To mitigate this, the proposed method applies a two-step planar approximation: global motion is first identified by performing a planar approximation over the entire sequence, and detailed planar approximation for each sub-interval is then conducted using only the inliers obtained from this global approximation. This ensures consistent tracking of the same object edge across different sub-intervals.
Furthermore, if the gradient information obtained from each interval is used directly, spatial misalignment occurs as the object moves over time as shown in Figure 4. Therefore, the edge position is determined based on the planar approximation result at a reference time, and the gradient information from each sub-interval is aggregated into this fixed spatial position. When computing the forward flow, the reference time is set to . Here, forward flow denotes the displacement field from the reference time to . This process compensates for spatial displacement over time, maintaining spatiotemporal consistency as the gradient information of the same object appears at identical coordinates across all intervals.

Figure 4. Spatial compensation of edge transitions of Figure 3. Gradients from sub-intervals (left) are spatially compensated to the reference position (right). This aggregates the spatiotemporal gradients of the same object at identical coordinates to maintain spatiotemporal consistency.
Formulation with Temporal Regularization
To estimate the flow, the obtained gradients are applied to an objective function. We modify (3) by adding a temporal regularization term multiplied by a weight :
The flow that minimizes the objective function (4) can be derived from the corresponding Euler–Lagrange equations, which, setting for simplicity, are given by:
By introducing the spatial neighborhood average and the temporal neighborhood average , (5) can be approximated as follows:
where is the weighted average of the 8 neighboring pixels at the same time [13], and is the average of the flow vectors at the same pixel between the preceding and succeeding intervals, following the spatial compensation described in Figure 4. The parameters and are coefficients associated with the space and time step sizes and , respectively. This discretization approach, which utilizes both spatial and temporal smoothness, can be viewed as an extension of the spatiotemporal Horn–Schunck framework [15, 16] applied to our event-based planar approximation.
To solve (6), the flow is obtained by iteratively applying the following update equations:
where and denote the computed results after the -th iteration.
Specifically, a straightforward pixel-wise linear regression is applied along the temporal axis to the time-series velocity fields to estimate the initial velocity field and the mean apparent acceleration over the interval . The forward flow is then obtained as the displacement from to :
Experiments
Experimental Setup
We evaluated the method on training sequences of the DSEC dataset [17], which contains real-world automotive data, by computing the flow for each 100 ms interval (i.e., ms), where ground-truth forward flow is available.
In the experiment, the VGA images (640 × 480 pixels) were divided into small regions of 40 × 40 pixels, and the two-step planar approximation described in Sec. IV was applied, with the number of temporal divisions set to . Specifically, using inliers extracted from the entire 100 ms interval, sub-intervals spanning a fraction of the total length were arranged with temporal overlap and a detailed planar approximation was performed. We set and empirically. Since the computational cost grows roughly linearly with whereas a larger is not expected to improve accuracy substantially, we considered sufficient. The parameter balances temporal resolution against data sparsity.
Then, from the spatiotemporal gradients obtained for each interval, the flow was computed by solving (6) under the spatiotemporal Neumann boundary condition. Notably, we replaced the standard RANSAC [19] used in the previous study [9] with MSAC (M-estimator Sample Consensus) [20, 21] and optimized some parameters. The inlier thresholds for the 1st and 2nd MSAC were set to 3.0 and 1.0, respectively. The objective function weights were set to and , where is relatively large to compensate for data sparsity, while is small but nonzero, balancing temporal stability against over-smoothing of motion dynamics. Optimization was performed for 5000 iterations initialized from zero velocity.
We used the Average Endpoint Error (AEE) [pixel] as the evaluation metric, defined as the norm of the error vector between the estimated flow and the ground-truth flow , i.e., , averaged over pixels
| Method | zurich_city_01_a | zurich_city_02_a | thun_00_a | |||||||||
| All | s-0-10 | s-10-40 | s-40+ | All | s-0-10 | s-10-40 | s-40+ | All | s-0-10 | s-10-40 | s-40+ | |
| Shiba et al. (Flow, L1) [8]1 | 6.17 | 3.16 | 10.64 | 41.76 | 7.60 | 5.75 | 10.57 | 35.81 | 3.73 | 3.02 | 4.96 | 27.40 |
| Frame Flow2 | 8.37 | 3.90 | 15.28 | 46.16 | 14.87 | 5.07 | 18.34 | 44.89 | 7.89 | 3.40 | 16.67 | 44.62 |
| Baseline [9] | 3.38 | 2.27 | 4.84 | 20.55 | 4.24 | 2.37 | 4.85 | 22.44 | 4.10 | 2.45 | 7.32 | 28.60 |
| +Time Div. | 3.14 | 2.24 | 4.24 | 19.43 | 3.97 | 2.32 | 4.38 | 21.23 | 3.53 | 2.19 | 6.16 | 27.15 |
| +Spatial Comp. | 3.06 | 2.29 | 3.93 | 19.66 | 3.96 | 2.27 | 4.26 | 21.83 | 3.40 | 2.20 | 5.74 | 26.93 |
| +Temporal Reg. (Proposed) | 3.07 | 2.31 | 3.94 | 19.27 | 3.99 | 2.30 | 4.29 | 21.71 | 3.37 | 2.20 | 5.63 | 26.20 |
Table 1. Optical Flow Estimation Results on the DSEC Dataset [17] (AEE [pixel] )
with available ground-truth. Ground-truth acceleration was approximated from temporal differences of adjacent ground-truth flows.
Evaluation Results and Ablation Analysis
Table I presents the optical flow estimation results on three DSEC [17] sequences. The proposed method surpasses the baseline [9] overall and in most motion ranges, and also shows competitive performance against [8]. Note that our implemented baseline already incorporates improvements over the original—replacing standard RANSAC with MSAC and optimizing parameters—so the baseline itself already surpasses the scores reported in [9].
The ablation study clarifies the role of each component. Temporal division (+Time Div.) consistently reduces AEE, with larger gains in moderate and high speed ranges, confirming that relaxing the constant-velocity assumption is broadly effective regardless of motion magnitude. Spatial compensation (+Spatial Comp.) contributes to improving accuracy primarily in moderate-motion regions (s-10-40). Furthermore, the addition of temporal regularization (+Temporal Reg.) is effective for high-motion regions (s-40+). By temporally stabilizing the velocity estimation for large motions, which inherently
tend to become unstable, it helps improve accuracy in the s-40+ range across all sequences. Meanwhile, the degradation in slow-motion regions (s-0-10) is relatively minor.
As visualized in Figure 5, the estimated forward flow is qualitatively consistent with the ground truth. Meanwhile, despite challenges such as event sparsity and the difficulty of local plane fitting, the proposed method outputs apparent acceleration estimates with discernible differences in color distributions among sequences, indicating that motion orientations are roughly captured. Further improving the accuracy of acceleration estimation remains future work.

Figure 5. Visual comparison of results. Rows from top to bottom correspond to the zurich_city_01_a, zurich_city_02_a, and thun_00_a sequences. Columns from left to right: Frame Image, Event Frame, Ground Truth (GT) Flow, GT Acceleration, Baseline Flow (without temporal division), Estimated Flow and Acceleration. Both estimated results are obtained by the proposed method with temporal divisions. The acceleration maps are shown only to illustrate the temporal variation captured by the proposed model; the primary evaluation criterion is the forward-flow AEE.
Conclusion
In this paper, we proposed an event-based optical flow estimation method that improves estimation accuracy by accounting for temporal velocity variations. By partitioning the space-time volume into sub-intervals and applying spatial compensation to maintain consistency, our method improved the accuracy over the conventional baseline. Temporal regularization further stabilizes the estimation, especially in high-motion regions, although its effect on AEE is sequence-dependent. While the computational complexity scales roughly linearly with the number of sub-intervals , reducing runtime through parallel processing and coarse-to-fine schemes remains an important direction for future work. Other future directions include comprehensive benchmarking against [8] and learning-based methods, and addressing event sparsity through more advanced surface approximation techniques.
References
References
- [1]G. Gallego, T. Delbrück, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis, and D. Scaramuzza, “Event-based vision: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 154–180, 2022. [Online]. Available: https://doi.org/10.1109/TPAMI.2020.3008413
- [2]D. Rossi, P. Vasseur, F. Morbidi, C. Demonceaux, and F. Rameau, “Event-aware distilled DETR for object detection in an automotive context,” in 2025 IEEE Intelligent Vehicles Symposium (IV), 2025, pp. 1849–1855. [Online]. Available: https://doi.org/10.1109/IV51958.2025.11097744
- [3]Y. Yang, J. Zhang, Y. Zhang, Y. Wei, D. Zou, J. S. Ren, and B. Shi, “Event-guided HDR reconstruction with diffusion priors,” in 2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 11 787–11 796. [Online]. Available: https://doi.org/10.1109/ICCV51701.2025.01096
- [4]A. Linares-Barranco, F. Perez-Peña, D. P. Moeys, F. Gomez-Rodriguez, G. Jimenez-Moreno, S.-C. Liu, and T. Delbruck, “Low latency event-based filtering and feature extraction for dynamic vision sensors in real-time FPGA applications,” IEEE Access, vol. 7, pp. 134 926–134 942, 2019. [Online]. Available: https://doi.org/10.1109/ACCESS.2019.2941282
- [5]I. Gokarn and A. Misra, “Poster: Profiling event vision processing on edge devices,” in Proceedings of the 22nd Annual International Conference on Mobile Systems, Applications and Services, ser. MOBISYS ’24. New York, NY, USA: Association for Computing Machinery, 2024, pp. 672–673. [Online]. Available: https://doi.org/10.1145/3643832.3661415
- [6]B. Li, J. Han, Y. Xu, and K. Rose, “Optical flow based co-located reference frame for video compression,” IEEE Transactions on Image Processing, vol. 29, pp. 8303–8315, 2020. [Online]. Available: https://doi.org/10.1109/TIP.2020.3014723
- [7]A. Agarwal, S. Gupta, and D. K. Singh, “Review of optical flow technique for moving object detection,” in 2016 2nd International Conference on Contemporary Computing and Informatics (IC3I), 2016, pp. 409–413. [Online]. Available: https://doi.org/10.1109/IC3I.2016.7917996
- [8]S. Shiba, Y. Klose, Y. Aoki, and G. Gallego, “Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 7742–7759, 2024. [Online]. Available: https://doi.org/10.1109/TPAMI.2024.3396116
- [9]T. Eda and Y. Kameda, “Optical flow constraint equation based on edge-line translation of event camera data,” Journal of Electronic Imaging, vol. 34, no. 4, p. 043024, 2025. [Online]. Available: https://doi.org/10.1117/1.JEI.34.4.043024
- [10]M. Koike, Y. Kameda, S. Sato, and T. Hamamoto, “Optical flow estimation via apparent acceleration from spatiotemporal plane fitting of event point clouds,” in Proceedings of the Picture Coding Symposium of Japan / Image Media Processing Symposium (PCSJ/IMPS), 2025, (in Japanese).
- [11]A. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “EV-FlowNet: Self-supervised optical flow estimation for event-based cameras,” in Proceedings of Robotics: Science and Systems, Pittsburgh, Pennsylvania, June 2018. [Online]. Available: https://doi.org/10.15607/RSS.2018.XIV.062
- [12]M. Gehrig, M. Millhäuser, D. Gehrig, and D. Scaramuzza, “E-RAFT: Dense optical flow from event cameras,” in 2021 International Conference on 3D Vision (3DV), 2021, pp. 197–206. [Online]. Available: https://doi.org/10.1109/3DV53792.2021.00030
- [13]B. K. Horn and B. G. Schunck, “Determining optical flow,” Artificial Intelligence, vol. 17, no. 1, pp. 185–203, 1981. [Online]. Available: https://doi.org/10.1016/0004-3702(81)90024-7
- [14]Y. Kameda, I. Matsuda, and S. Itoh, “Numerically stable multi-channel depth scene flow with adaptive weighting of regularization terms,” in 2020 28th European Signal Processing Conference (EUSIPCO), 2021, pp. 605–609. [Online]. Available: https://doi.org/10.23919/Eusipco47968.2020.9287861
- [15]H.-H. Nagel, “Extending the ‘Oriented smoothness constraint’ into the temporal domain and the estimation of derivatives of optical flow,” in Computer Vision – ECCV 90, O. Faugeras, Ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 1990, pp. 139–148. [Online]. Available: https://doi.org/10.1007/BFb0014860
- [16]J. Weickert and C. Schnörr, “Variational optic flow computation with a spatio-temporal smoothness constraint,” Journal of Mathematical Imaging and Vision, vol. 14, no. 3, pp. 245–255, May 2001. [Online]. Available: https://doi.org/10.1023/A:1011286029287
- [17]M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “DSEC: A stereo event camera dataset for driving scenarios,” IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4947–4954, 2021. [Online]. Available: https://doi.org/10.1109/LRA.2021.3068942
- [18]S. Shiba, “GitHub - tub-rip/event_based_optical_flow: The official implementation of ‘Secrets of Event-based Optical Flow’ (ECCV 2022 oral and IEEE T-PAMI 2024).” [Online]. Available: https://github.com/tub-rip/event_based_optical_flow
- [19]M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395, Jun. 1981. [Online]. Available: https://doi.org/10.1145/358669.358692
- [20]P. Torr and A. Zisserman, “Robust computation and parametrization of multiple view relations,” in Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), 1998, pp. 727–732. [Online]. Available: https://doi.org/10.1109/ICCV.1998.710798
- [21]———, “MLESAC: A new robust estimator with application to estimating image geometry,” Computer Vision and Image Understanding, vol. 78, no. 1, pp. 138–156, 2000. [Online]. Available: https://doi.org/10.1006/cviu.1999.0832
