Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Event3R: Asynchronous-to-Global 3D Reconstruction from Event Camera via Spatial-Temporal Feature Aggregation

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Event3R claims that asynchronous event streams alone, passed through a temporal-attention voxel network, directly produce globally aligned 3D point clouds without RGB images, optical flow, or iterative optimization.

desk verdict Event3R is a plausible event-only adaptation of DUSt3R with a useful temporal encoder and self-supervised trick, but the feed-forward claim is contradicted by its own supplement for N>2, and the evaluation is partly circular. Worth engaging, needs revision. read the letter →

arxiv 2607.15727 v1 pith:L4LBC3TB submitted 2026-07-17 cs.CV

classification cs.CV MSC 68T4568T07
keywords eventcamera3Dreconstructionfeed-forwardnetworkspatial-temporalvoxelstemporalattentionself-supervisedpretrainingdepthestimationpose
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that event cameras—sensors that fire asynchronous brightness-change pixels—can drive the same kind of instant, globally consistent 3D reconstruction that image-based feed-forward networks achieve. It argues that by voxelizing event streams into time bins and letting a temporal-attention module aggregate features across bins, a network can consume two or more short event segments and output camera-aligned dense geometry in roughly 0.2 seconds. The payoff: robots and AR systems could reconstruct scenes under fast motion or extreme lighting where frame-based cameras blur or saturate. The authors claim large gains over existing event-based depth, pose, and reconstruction baselines, and show that self-supervised bin-masking pretraining reduces dependence on labeled event data.

What carries the argument

The load-bearing mechanism is the spatial-temporal voxel representation plus temporal attention: each event is splatted into a H×W×T voxel grid with trilinear interpolation, and the temporal encoder runs self-attention over the T bin embeddings at each spatial patch, aggregating motion cues into the central bin's feature. The aggregated features are consumed by an RGB-pretrained geometric backbone that predicts dense depth and camera-aligned pointmaps. Masked Bin Modeling is the self-supervised companion: it masks whole temporal bins and trains a lightweight decoder to reconstruct them, enforcing inter-bin temporal modeling.

What would settle it

Train the same architecture from scratch without RGB-pretrained initialization and without synthetic events, using only real event data and unlabeled pretraining; if depth and pose accuracy collapse to baseline levels, the reported gains come from the inherited RGB priors rather than from event-specific temporal modeling. Alternatively, evaluate on a real event sequence with fast ego-motion and low event rate per bin and check whether temporal attention still improves over feeding raw voxels directly.

Watch

Extended reading notes

Core claim

The central claim is that asynchronous event streams can be treated as structured spatial-temporal voxel tensors and mapped end-to-end to a globally aligned 3D pointmap. A patch-level temporal encoder performs self-attention across voxelized time bins, treating the center bin as the temporal reference frame; the resulting features feed a spatial encoder and decoder inherited from an RGB-pretrained feed-forward reconstruction backbone. The authors add a Masked Bin Modeling objective—masking entire time bins and reconstructing their features from neighbors—used first for self-supervised pretraining and retained during fine-tuning, alongside contrastive and consistency losses. On synthetic and

Load-bearing premise

The spatial encoder and decoder are initialized from weights pretrained on ordinary RGB images, and most training events are simulated from RGB video, so the method assumes that geometry priors learned from images transfer to temporally aggregated event-voxel features and that synthetic events resemble real ones well enough.

Editorial extensions

If this is right

  • Feed-forward event-only reconstruction closes the loop for event cameras: no pose optimization, no RGB, no iterative rendering, usable at roughly 0.2 seconds per frame pair.
  • Temporal attention over voxel bins is the main accuracy lever: removing it roughly doubles depth error and triples trajectory error in the reported ablations.
  • Masked Bin Modeling pretraining substitutes for labeled data: disabling it degrades results substantially, so unlabeled event streams can build temporal priors.
  • The same pointmap output plugs into existing downstream pipelines for depth, pose, and multi-view fusion without modification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If RGB-pretrained geometry weights transfer to aggregated event-voxel features, then event voxels may occupy a representation space close enough to image patches that other feed-forward geometric models could be adapted with the same recipe, not just the one used here.
  • A testable extension: vary the number of temporal bins and event rate; the method's reliance on a central reference bin suggests accuracy should degrade asymmetrically for long windows where motion dominates a single bin.
  • The self-supervised masked-bin objective should in principle work with real unlabeled events alone; a clean experiment would pretrain on real event data with no simulated events and measure how much of the gain remains.
  • If the synthetic-to-real gap matters, releasing a real-event fine-tuning set would be the direct stress test; current real-data evaluation is limited to specific outdoor and indoor recordings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Event3R, a feed-forward, event-only framework for 3D reconstruction from asynchronous event streams. Events are converted into spatial-temporal voxels with T time bins; a temporal encoder performs self-attention across bins to aggregate motion cues, and a DUSt3R-initialized spatial encoder/decoder predicts dense, camera-aligned pointmaps. The authors introduce Masked Bin Modeling (MBM) for self-supervised pre-training and as an auxiliary fine-tuning objective, together with contrastive and consistency losses. The method is evaluated on depth estimation, camera pose estimation, and 3D reconstruction using TartanAir and MVSEC, reporting large improvements over event-based baselines.

Significance. If the claims are fully substantiated, Event3R would be a notable step toward bringing feed-forward, optimization-free 3D reconstruction to event cameras, with practical relevance for robotics and high-speed perception. The architecture is clearly described, and the ablations support the value of the temporal encoder and MBM within the proposed framework: removing them degrades depth and pose metrics substantially. The paper also makes an effort to compare with multiple event-based baselines. However, the central 'feed-forward, no iterative optimization' claim is internally inconsistent with the supplementary multi-view processing, and the evaluation protocol appears to fine-tune and evaluate on the same datasets without documented held-out splits. These two issues are load-bearing for the headline claims, so the current evidence is suggestive rather than conclusive.

major comments (3)
  1. [Section III and Supplementary Sec. VI.b] The paper states that Event3R 'produces a globally aligned 3D pointmap ... without relying on iterative optimization or external pose estimation' (Section III). However, Supplementary Sec. VI.b says that for N>2 views the authors 'follow the global alignment framework of DUSt3R' and fuse pairwise pointmaps 'through a global optimization that minimizes the weighted 3D reprojection error.' This is exactly an iterative optimization step. As written, the feed-forward/no-optimization claim is false for multi-view processing, and the global-consistency results in Fig. 5 and Table III are not produced by a single forward pass of the network. Please either restrict the claim to the two-view case, or report multi-view reconstruction with the network alone and document the optimization step separately.
  2. [Section IV.A.b, Tables I-III] The network is jointly fine-tuned on TartanAir and MVSEC ('In the second stage, the network undergoes joint training on the TartanAir and MVSEC datasets'), and the headline depth, pose, and 3D-reconstruction results are then reported on the same two datasets. No training/test split is documented. As written, this is a circular evaluation: the model may be memorizing training scenes rather than generalizing. This is particularly concerning for the MVSEC real-world results, since real event data are scarce and the paper claims real-world applicability. Please state which sequences were used for training and which for evaluation, and if standard splits exist, cite them explicitly; otherwise report a held-out evaluation.
  3. [Section IV.A.a and IV.A.b] The 'event-only' claim is weakened by the initialization and training-data choices: the spatial encoder and decoder are initialized from RGB-pretrained DUSt3R weights, and TartanAir event streams are simulated from RGB sequences using ESIM. Thus the method assumes that RGB-geometry priors transfer to temporally aggregated event features and that synthetic events are representative of real event streams. These are plausible assumptions, but they are load-bearing for the generalization claims. A controlled experiment (e.g., training without DUSt3R initialization, or evaluating on a real-event dataset not used in any training stage) would substantially strengthen the paper. As written, the gains could partly reflect RGB-prior reuse and synthetic-domain overfitting rather than a robust event-based capability.
minor comments (5)
  1. [Section IV.A.a] Equation (11) includes a weight α for the auxiliary MBM loss, but the implementation details never state the value of α. Please report it.
  2. [General] No statement of code, data, or checkpoint availability is included. Given the paper's claims and the field's reproducibility norms, please add an availability statement.
  3. [Table IV] The ablation table reports ATE but does not specify which scene/sequence is used. The full-model ATE of 0.1524 differs from all five sequences in Table II, so the reader cannot compare. Please state the evaluation protocol for the ablation.
  4. [Equation (1)] The normalized temporal coordinate is written as t*_k = T-1/Δ(t_k - t_0), which is ambiguous: it should presumably be (T-1)(t_k - t_0)/Δ. Please fix the notation.
  5. [Figure 4] The qualitative depth comparisons label rows only as 'Synthetic' or 'Real-world'; please identify the specific datasets/scenes so the reader can judge the difficulty of the examples.

Circularity Check

2 steps flagged · score 4.0 of 10

Multi-view 'feed-forward, no-optimization' claim contradicts the supplement's DUSt3R-style global optimization; headline benchmarks overlap the fine-tuning datasets.

  1. other [Section III (Method) vs. Supplementary Sec. VI.b (Multi-view Processing)]
    "Our method performs feed-forward 3D reconstruction directly from event streams. It takes two or multiple ( N >2) asynchronous event sequences as input and produces a globally aligned 3D pointmap within one second, without relying on iterative optimization or external pose estimation. ... To reconstruct a consistent 3D scene from N views using only event voxel grids, we follow the global alignment framework of DUSt3R... pairwise pointmaps with confidence maps, which are then fused through a global optimization that minimizes the weighted 3D reprojection error."

    The paper's central claim of feed-forward, optimization-free global reconstruction is contradicted by its own multi-view protocol. For N>2, the reported globally aligned pointmap is not the direct network output but the result of an explicit DUSt3R global optimization. Thus the headline multi-view result is produced by an imported optimization stage whose presence the main text denies, reducing the claimed feed-forward novelty for multi-view settings to a DUSt3R alignment post-process rather than an independent forward-pass prediction.

  2. fitted input called prediction [Section IV.A.b (Datasets); Tables I, II, III]
    "In the second stage, the network undergoes joint training on the TartanAir and MVSEC datasets. ... We evaluate monocular depth estimation on the TartanAirEvent and MVSEC datasets using the same standard metrics as DUSt3R and CUT3R."

    The same two datasets are used for fine-tuning and for the headline depth, pose, and reconstruction evaluations, with no described held-out split. The reported superiority over baselines is therefore at least partly in-sample on the training distribution; the 'predictions' in Tables I-III are measurements of the fitted model on data it was trained on. Unless a separate test split is explicitly defined, this makes the generalization claims statistically forced by construction.

full rationale

Event3R's core architecture is a standard supervised pipeline: external DUSt3R weights initialize the spatial encoder/decoder, ESIM-generated events from RGB sequences provide training data, and the temporal encoder plus MBM are auxiliary components that are individually ablated. I found no self-definitional equation, no imported uniqueness theorem, and no load-bearing self-citation chain; the IncEventGS reference is used as a comparison baseline, not as justification for the method's central claim. The main circularity-adjacent problems are the internal contradiction between the claimed 'feed-forward / no iterative optimization' behavior and the supplement's DUSt3R global-optimization procedure for N>2, and the evaluation protocol that fine-tunes and reports headline results on the same TartanAir/MVSEC datasets without a stated held-out split. These issues undermine the strength and interpretation of the claims, but the derivation itself is not equivalent to its inputs by definition.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on several hand-chosen hyperparameters, on simulated event fidelity, and on transfer from an RGB-pretrained backbone. No new physical entities are introduced. The key unverified assumptions are the synthetic-to-real transfer of ESIM events and the adequacy of the 50 ms / T=5 voxelization for global geometry.

free parameters (5)
  • Number of temporal bins T = T=5
    Chosen after balancing performance and efficiency (Sec IV.A.a), not derived from a physical or algorithmic constraint.
  • Event chunk length = 50 ms
    Fixed hyperparameter controlling how much event history is aggregated (Sec IV.A.a).
  • Auxiliary loss weights λ1, λ2 = λ1=1.0, λ2=1.0
    Set by hand for contrastive and consistency losses (Sec IV.A.a).
  • MBM auxiliary weight α = not specified
    Appears in L_joint (Eq. 11) but no value or selection procedure is reported.
  • Temporal encoder capacity = 2 layers, 4 heads, patch 8×8, dim 32
    Architecture choices made for the empirical setup; not justified by a theorem or systematic sweep.
assumptions (4)
  • domain assumption RGB-pretrained DUSt3R spatial encoder/decoder weights transfer to event-voxel-derived features.
    The network is initialized with DUSt3R weights and fine-tuned on event data; the paper provides no analysis of how much of the final performance comes from this RGB-geometry prior.
  • domain assumption ESIM-simulated event streams from RGB sequences are representative enough of real event statistics.
    MatrixCity and TartanAir events are generated by ESIM (Sec IV.A.b); the real MVSEC dataset is also used, but the heavy synthetic component is assumed to transfer.
  • domain assumption A short voxel window split into T=5 temporal bins with a center bin captures the motion cues needed for global alignment.
    The design sets T odd to define a central reference bin (Sec III.B); no ablation varying T or chunk length is reported.
  • domain assumption Self-supervised masked-bin pretraining on unlabeled event data improves downstream supervised reconstruction.
    The ablation shows a benefit when pretraining is enabled, but the mechanism is assumed rather than derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Event3R: Asynchronous-to-Global 3D Reconstruction from Event Camera via Spatial-Temporal Feature Aggregation." pith.science (2026). https://pith.science/paper/L4LBC3TB

@misc{pith2026260715727,
  author       = {Pith},
  title        = {Pith review of: Event3R: Asynchronous-to-Global 3D Reconstruction from Event Camera via Spatial-Temporal Feature Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L4LBC3TB}},
  note         = {Machine review of arXiv:2607.15727}
}
read the original abstract

Robust 3D reconstruction is essential for robotics and embodied perception. Recent feed-forward approaches such as DUSt3R have demonstrated impressive progress in dense 3D reconstruction from RGB images, achieving global geometric consistency and strong generalization. However, extending such dense 3D reconstruction to event cameras remains challenging due to their asynchronous, sparse, and highly dynamic nature, as well as the lack of large-scale, well-labeled datasets. In this work, we introduce Event3R, a feed-forward framework that directly maps asynchronous event streams to globally consistent 3D point clouds. Event3R represents incoming events as spatial-temporal voxels, enabling time-aware feature integration through a temporal attention module that enhances the module's temporal feature learning. To further strengthen temporal representation learning and reduce reliance on labeled data, we propose a Masked Bin Modeling (MBM) strategy for self-supervised pre-training, enabling robust temporal representation learning with minimal labeled data, and retain it as an auxiliary fine-tuning objective. In addition, contrastive alignment and consistency regularization losses are incorporated during fine-tuning to reinforce structural correspondence and temporal coherence across views. Extensive experiments on both synthetic and real-world benchmarks demonstrate that Event3R achieves robust, temporally consistent, and globally aligned 3D reconstructions, significantly outperforming existing event-based methods.

Figures

Figures reproduced from arXiv: 2607.15727 by the authors.

Figure 1
Figure 1. Illustration of our Event3R. Event3R takes two or more short segments of asynchronous event streams, converts them into spatial–temporal voxel representations, and directly predicts a globally aligned 3D pointmap in about 0.2s. For visualization, the pointcloud is colored using the RGB image corresponding to the center voxel bin, although no RGB data is used during training or inference. Abstract— Robust 3D reconstr… view at source ↗
Figure 2
Figure 2. Qualitative depth estimation comparison with DUSt3R under challenging conditions. Event3R is evaluated in a zero-shot setting without training on the PEOD [9] dataset. event–depth–pose datasets. To fill this research gap, we present Event3R, a feed￾forward framework designed to bring global 3D reconstruction to asynchronous event data. Our key insight is to reinterpret event streams as structured spatial–temporal vo… view at source ↗
Figure 3
Figure 3. Overview of the Event3R pipeline. (a) Full pipeline: asynchronous event streams are converted into spatial–temporal voxels (T bins) and processed by the temporal encoder to capture motion cues. The resulting features are then passed through a spatial encoder and decoder, built on a DUSt3R backbone, to predict dense, globally aligned 3D pointmaps. (b) Temporal Encoder: self-attention across voxelized time bins enable… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of depth estimation results on both synthetic and real-world datasets. For better visualization, the grayscale depth maps are colorized using the ’jet’ colormap. tency losses—on depth and pose estimation. Removing the Temporal Encoder (i.e., feed…
Figure 5
Figure 5. Figure 5: Global multi-view 3D reconstruction results. We demonstrate the capability of Event3R to reconstruct globally consistent 3D scenes from multiple views (N = 5). The estimated camera trajectories are visualized as frustums alongside the point clouds, demonstrating accura…
Figure 6
Figure 6. Figure 6: Qualitative results of two-view 3D reconstruction on synthetic datasets. We visualize the reconstructed point clouds from two different viewpoints (View 1 and View 2). For visualization purposes, the point clouds are colored using the ground truth RGB images [PITH_FUL…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

  1. [1]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tanciket al., “NeRF: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

  2. [2]

    3D Gaussian Splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhleret al., “3D Gaussian Splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 1–14, 2023

  3. [3]

    Dust3R: Geometric 3D vision made easy,

    S. Wang, V . Leroy, Y . Cabonet al., “Dust3R: Geometric 3D vision made easy,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 697–20 709

  4. [4]

    Continuous 3D perception model with persistent state,

    Q. Wang, Y . Zhang, A. Holynskiet al., “Continuous 3D perception model with persistent state,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 10 510–10 522

  5. [5]

    VGGT: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaevet al., “VGGT: Visual geometry grounded transformer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 5294–5306

  6. [6]

    EvGGS: A collaborative learning framework for event-based generalizable Gaussian Splatting,

    J. Wang, J. He, Z. Zhanget al., “EvGGS: A collaborative learning framework for event-based generalizable Gaussian Splatting,”arXiv preprint arXiv:2405.14959, 2024

  7. [7]

    Event3dgs: Event-based 3d gaussian splatting for high- speed robot egomotion,

    T. Xiong, J. Wu, B. He, C. Fermuller, Y . Aloimonos, H. Huang, and C. Metzler, “Event3dgs: Event-based 3d gaussian splatting for high- speed robot egomotion,” in8th Annual Conference on Robot Learning

  8. [8]

    IncEventGS: Pose-free Gaussian Splatting from a single event camera,

    J. Huang, C. Dong, X. Chenet al., “IncEventGS: Pose-free Gaussian Splatting from a single event camera,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 26 933–26 942

Show all 35 references
  1. [9]

    PEOD: A pixel-aligned event-rgb benchmark for object detection under challenging conditions,

    L. Cui, H. Liu, M. Liuet al., “PEOD: A pixel-aligned event-rgb benchmark for object detection under challenging conditions,”arXiv preprint arXiv:2511.08140, 2025

  2. [10]

    Structure-From-Motion revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-From-Motion revisited,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4104–4113

  3. [11]

    CREPES: Cooperative RElative pose estimation system,

    Z. Xun, J. Huang, Z. Liet al., “CREPES: Cooperative RElative pose estimation system,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 5274–5281

  4. [12]

    Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tanciket al., “Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 5855–5864

  5. [13]

    COLMAP-free 3D Gaussian Splatting,

    Y . Fu, S. Liu, A. Kulkarniet al., “COLMAP-free 3D Gaussian Splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 796–20 805

  6. [14]

    Grounding image matching in 3D with MASt3R,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3D with MASt3R,” inEuropean Conference on Computer Vision (ECCV), 2024, pp. 71–91

  7. [15]

    Review of Feed-Forward 3D recon- struction: From DUSt3R to VGGT,

    W. Zhang, Y . Wu, S. Liet al., “Review of Feed-Forward 3D recon- struction: From DUSt3R to VGGT,”arXiv preprint arXiv:2507.08448, 2025

  8. [16]

    Real-time 3D recon- struction and 6-DoF tracking with an event camera,

    H. Kim, S. Leutenegger, and A. J. Davison, “Real-time 3D recon- struction and 6-DoF tracking with an event camera,” inEuropean Conference on Computer Vision (ECCV), 2016, pp. 349–364

  9. [17]

    EMVS: Event-based multi-view stereo—3D reconstruction with an event camera in real- time,

    H. Rebecq, G. Gallego, E. Mueggleret al., “EMVS: Event-based multi-view stereo—3D reconstruction with an event camera in real- time,”International Journal of Computer Vision, vol. 126, no. 12, pp. 1394–1414, 2018

  10. [18]

    EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real time,

    H. Rebecq, T. Horstschaefer, G. Gallegoet al., “EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real time,”IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 593– 600, 2017

  11. [19]

    Semi-dense 3D reconstruction with a stereo event camera,

    Y . Zhou, G. Gallego, H. Rebecqet al., “Semi-dense 3D reconstruction with a stereo event camera,” inEuropean Conference on Computer Vision (ECCV), 2018, pp. 235–251

  12. [20]

    E-NeRF: Neural radiance fields from a moving event camera,

    S. Klenk, L. Koestler, D. Scaramuzzaet al., “E-NeRF: Neural radiance fields from a moving event camera,”IEEE Robotics and Automation Letters, vol. 8, no. 3, pp. 1587–1594, 2023

  13. [21]

    EV-NeRF: Event based neural radiance field,

    I. Hwang, J. Kim, and Y . M. Kim, “EV-NeRF: Event based neural radiance field,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023, pp. 837–847

  14. [22]

    EventNeRF: Neural radiance fields from a single colour event camera,

    V . Rudnev, M. Elgharib, C. Theobaltet al., “EventNeRF: Neural radiance fields from a single colour event camera,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4992–5002

  15. [23]

    EventSplat: 3D Gaussian Splatting from moving event cameras for real-time rendering,

    T. Yura, A. Mirzaei, and I. Gilitschenski, “EventSplat: 3D Gaussian Splatting from moving event cameras for real-time rendering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 26 876–26 886

  16. [24]

    Event-3DGS: Event-based 3D reconstruc- tion using 3D Gaussian Splatting,

    H. Han, J. Li, H. Weiet al., “Event-3DGS: Event-based 3D reconstruc- tion using 3D Gaussian Splatting,”Advances in Neural Information Processing Systems (NeurIPS), vol. 37, pp. 128 139–128 159, 2024

  17. [25]

    EAG3R: Event-augmented 3D geometry estimation for dynamic and extreme-lighting scenes,

    X. Wu, Y . Yu, X. Lyuet al., “EAG3R: Event-augmented 3D geometry estimation for dynamic and extreme-lighting scenes,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025

  18. [26]

    Masked event modeling: Self-supervised pretraining for event cameras,

    S. Klenk, D. Bonello, L. Koestleret al., “Masked event modeling: Self-supervised pretraining for event cameras,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024, pp. 2378–2388

  19. [27]

    MatrixCity: A large-scale city dataset for city-scale neural rendering and beyond,

    Y . Li, L. Jiang, L. Xuet al., “MatrixCity: A large-scale city dataset for city-scale neural rendering and beyond,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 3205–3215

  20. [28]

    TUM-VIE: The TUM stereo visual-inertial event dataset,

    S. Klenk, J. Chui, N. Demmelet al., “TUM-VIE: The TUM stereo visual-inertial event dataset,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 8601–8608

  21. [29]

    Video to events: Recycling video datasets for event cameras,

    D. Gehrig, M. Gehrig, J. Hidalgo-Carri ´oet al., “Video to events: Recycling video datasets for event cameras,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

  22. [30]

    The MultiVehicle stereo event camera dataset: An event camera dataset for 3D perception,

    A. Z. Zhu, D. Thakur, T. ¨Ozaslanet al., “The MultiVehicle stereo event camera dataset: An event camera dataset for 3D perception,”IEEE Robotics and Automation Letters, vol. 3, no. 3, pp. 2032–2039, 2018

  23. [31]

    Learning monocular dense depth from events,

    J. Hidalgo-Carri´o, D. Gehrig, and D. Scaramuzza, “Learning monocular dense depth from events,” inInternational Conference on 3D Vision (3DV), 2020, pp. 534–542

  24. [32]

    DERD-Net: Learning depth from event-based ray densities,

    D. Hitzges, S. Ghosh, and G. Gallego, “DERD-Net: Learning depth from event-based ray densities,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025

  25. [33]

    Event-based monocular dense depth esti- mation with recurrent transformers,

    X. Liu, J. Li, X. Fanet al., “Event-based monocular dense depth esti- mation with recurrent transformers,”arXiv preprint arXiv:2212.02791, 2022

  26. [34]

    Deep event visual odometry,

    S. Klenk, M. Motzet, L. Koestleret al., “Deep event visual odometry,” inInternational Conference on 3D Vision (3DV), 2024, pp. 739–749

  27. [35]

    Events-to-Video: Bringing modern computer vision to event cameras,

    H. Rebecq, R. Ranftl, V . Koltunet al., “Events-to-Video: Bringing modern computer vision to event cameras,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. Event3R: Asynchronous-to-Global 3D Reconstruction from Event Camera vi...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.