eBIRD: Event-based Intensity Image Reconstruction Using Controllable Diffusion Models

Fabian Valderrama; Ignacio Bugueno-Cordova; Rodrigo Verschae

REVIEW 3 major objections 8 minor 23 references

eBIRD: Event-based Intensity Image Reconstruction Using Controllable Diffusion Models

T0 review · 3 major / 8 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that event-camera streams can be converted to intensity images by a diffusion model conditioned through ControlNet on event frames, and that whether generic or specialized training works better depends on the variability…

desk verdict A workmanlike DDPM+ControlNet paper whose only genuinely new contribution is a generic-vs-specialized training comparison, and that comparison is not yet trustworthy on the face side because the fine-tuning split may overlap with the test subjects. read the letter →

arxiv 2608.08519 v1 pith:G5JVIOTY submitted 2026-08-09 cs.CV

Ignacio Bugueno-Cordova , Fabian Valderrama , Rodrigo Verschae This is my paper · ORCID

classification cs.CV

keywords eventcamerasimagereconstructiondiffusionmodelsControlNetconditioningframesneuromorphicvisionclass-agnosticlearningclass-specific

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

This paper proposes eBIRD, a two-stage framework that first trains a denoising diffusion backbone on ordinary images and then freezes it while a ControlNet branch learns to condition the denoising process on 33 ms event frames, so the network maps event streams to intensity images. It reports that on N-MNIST the generic model eBIRD-G beats ten class-specific models (MSE 0.0052 vs 0.0074, SSIM 0.8982 vs 0.8691, PSNR 23.34 vs 21.88 dB), while on RGBE-Gaze the user-specialized model eBIRD-S beats both the generic variant and the Pix2Pix, E2VID, and ET-Net baselines (MSE 0.0161, SSIM 0.7615, PSNR 19.08 dB). The intended upshot is that class-agnostic learning is preferable where appearance variability is low, and class- or subject-specific adaptation is preferable where variability is high.

What carries the argument

The mechanism is a two-stage diffusion pipeline. Stage one trains a standard DDPM U-Net to denoise target images unconditionally; stage two freezes that backbone and trains a ControlNet branch, of identical topology and connected via zero-initialized convolutions, to inject features derived from an event frame at multiple resolutions, with the training objective predicting the noise added at each diffusion timestep. The conditioning signal is a simple event frame, the per-pixel accumulation of positive and negative event polarities within a 33 ms window. This setup separates the image prior from the event-to-image mapping, allowing the same prior to be specialized by task.

What would settle it

Evaluate eBIRD-S on subjects excluded from its fine-tuning set while keeping all other training conditions fixed. If the held-out-subject performance drops to eBIRD-G levels (or worse), the specialization advantage on faces is memorization, not generalization; if it persists, the paper's conclusion survives.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that event-guided intensity reconstruction can be cast as a controllable diffusion problem: a pretrained DDPM U-Net, kept frozen, is steered by a ControlNet branch whose only input is an event frame, and this conditioned denoiser reconstructs recognizable handwritten digits and faces from sparse, asynchronous event data. The second discovery is a domain-dependence in learning strategy: a single model trained on all digit classes outperforms independent per-class models, whereas a model fine-tuned on the target users outperforms the generic model on faces. The authors interpret this as evidence that specialization is not a property of the architecture but of the reconstruction domain's variability.

Load-bearing premise

The face-specialization result assumes the 'target users' used to fine-tune eBIRD-S are not the same people later used for evaluation; otherwise the reported gains reflect memorization of identities rather than generalization from events to images.

Editorial extensions

If this is right

For low-variability domains like handwritten digits, practitioners should prefer a single shared diffusion prior over per-class models: it yields lower MSE and higher SSIM/PSNR on every digit class.
For high-variability domains like faces, fine-tuning the conditioned model on the target users is the better route, and the resulting reconstructions beat Pix2Pix and the recurrent E2VID and ET-Net baselines by large margins.
The two-stage recipe—unconditional image prior, then frozen-backbone ControlNet conditioning—can be reused for other event-based reconstruction tasks without retraining the generative backbone.
Because the same framework performs well on both 28×28 digits and full-face images, the approach scales beyond small synthetic neuromorphic datasets to realistic event-camera data.

Reading between the lines

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

Inference: The paper does not report whether the RGBE-Gaze evaluation subjects were excluded from eBIRD-S's fine-tuning set; if they were not, the specialization gain could partly be identity memorization rather than better event-to-image mapping.
Inference: The E2VID and ET-Net baselines are applied with pretrained weights trained on other event domains, so their gap to eBIRD may shrink if those models were fine-tuned on RGBE-Gaze; a fair comparison would require retraining them under the same protocol.
Inference: The domain-dependent conclusion suggests a practical rule: measure intra-class or intra-subject variability before choosing between shared and specialized diffusion models for event reconstruction.
Inference: The 33 ms event-frame representation discards temporal ordering within the window; using time surfaces or voxel grids as conditioning might improve fidelity, which the paper lists as future work.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is a straightforward DDPM+ControlNet application to event-to-image reconstruction. The framework itself is not new — that combination is already in the literature — and the authors are honest about that. What is new is the systematic comparison between a single generic model and per-class/per-subject specialized models, with a clean story: generic wins on digits, specialized wins on faces. If that result holds up, it is a useful practical rule of thumb for the subfield.

The N-MNIST experiment is the strongest part. The protocol is clear: ten per-class models and one general model, all trained from scratch on the same event representation, evaluated on the standard test split. The numbers consistently favor the general model, and the qualitative examples support it. That half of the paper is solid.

The RGBE-Gaze half is where it gets shaky. Section 4.3 says eBIRD-S is obtained by fine-tuning eBIRD-G "using data from the target users," and the stress-test raises the obvious question: are those users the same people used in the evaluation? The paper says all methods used "the same train, validation and test split," but it never states the split is subject-disjoint. If the same faces appear in fine-tuning and test, then eBIRD-S is partly measuring identity memorization, not event-to-image generalization. The qualitative figures are equally consistent with that reading. This is a load-bearing ambiguity and it needs to be fixed before the central claim is credible.

There are also smaller problems. E2VID and ET-Net are evaluated with frozen pretrained weights on a domain they never trained on, so the headline comparison against them is not apples-to-apples. N-MNIST results have no error bars. There is no ablation, and the paper explicitly defers it. And since the authors cite two prior diffusion-based event reconstruction works, not comparing against them is a noticeable gap. The citations themselves look appropriate and honest.

Who is this for? People working on event camera reconstruction who want a quick check on whether diffusion priors plus ControlNet are worth trying, and whether to bother training per-domain models. It is a preliminary empirical note rather than a definitive study, but the question is real and the flaws are fixable.

My recommendation: do not desk-reject, but do send it to reviewers with the expectation of serious revision. The subject-split issue alone is enough to require re-running the RGBE-Gaze experiments with subject-disjoint splits and reporting the numbers. With that, the paper could make a modest but genuine contribution.

Referee Report

3 major / 8 minor

Summary. The paper proposes eBIRD, a two-stage framework that first trains a DDPM on intensity images and then freezes the backbone while training a ControlNet branch conditioned on event frames. It compares a generic model (eBIRD-G) trained on all training samples with specialized models (eBIRD-S) trained per digit class on N-MNIST or fine-tuned on target users of RGBE-Gaze. The authors report that eBIRD-G outperforms specialized models on digits (MSE 0.0052 vs 0.0074) while eBIRD-S outperforms eBIRD-G on faces (MSE 0.0161 vs 0.0276), leading to the conclusion that the preferred learning strategy depends on domain variability.

Significance. If the comparisons were clean, the paper would provide a useful data point on when to specialize conditional diffusion models for event-based reconstruction, and the proposed framework is a sensible combination of established components (DDPM plus ControlNet). The two-stage training is clearly described, and the N-MNIST comparison appears internally consistent. However, the face-based half of the central claim is currently supported by a protocol that may allow identity memorization, and the state-of-the-art comparison is confounded by out-of-domain pretrained baselines. The paper also omits ablations and error bars. Therefore, the significance is conditional on fixing the evaluation protocol.

major comments (3)

[Section 5.2 and Table 2] The fine-tuning data for eBIRD-S on RGBE-Gaze consists of "data from the target users" without stating that these users are disjoint from the evaluation subjects. If the same subjects appear in training and evaluation, the eBIRD-S advantage (MSE 0.0161 vs 0.0276, SSIM 0.7615 vs 0.6381, PSNR 19.08 vs 16.15 dB) can be explained by the ControlNet branch learning identity-specific appearance rather than by a genuine advantage of specialization for face reconstruction. The manuscript must state the subject-level split and, ideally, evaluate on held-out subjects that were never used for fine-tuning. This is load-bearing for the conclusion in Section 7 that specialization is superior for faces.
[Section 5.2 and Table 2] E2VID and ET-Net are evaluated with pretrained weights without fine-tuning on RGBE-Gaze. Those models were trained on different event camera data or different event representations, so their poor performance (MSE around 0.10, PSNR around 10 dB) may simply reflect domain shift rather than architectural inferiority. The claim that eBIRD "substantially improves" over these methods is not supported unless the baselines are trained or at least adapted to the same event representation, resolution, and domain. This comment does not affect the eBIRD-G versus eBIRD-S comparison, but it weakens the state-of-the-art positioning.
[Section 6.1 and Table 1] No standard deviations or significance tests are reported for the N-MNIST results. The differences between the generic and class-specific models, while consistent across classes, are small (for example, average MSE 0.0052 versus 0.0074). Without repeated runs or statistical analysis, it is unclear whether the difference is meaningful. This is relevant because the N-MNIST experiment is the cleaner of the two comparisons.

minor comments (8)

[Abstract and Table 2] The abstract reports an SSIM of 0.7605 for eBIRD-S, while Table 2 reports 0.7615; please correct the inconsistency.
[Section 2] The phrase "we study of class-agnostic" should be "we study the effect of class-agnostic" or similar; the current wording is ungrammatical.
[Section 6.2] The section heading "Face Reconstruction using RGB-E Gaze" is inconsistent with "RGBE-Gaze" used elsewhere; please standardize the dataset name.
[Table 1] The column headers S1 through S0 and G1 through G0 are ambiguous; please label them by digit class (0-9) for clarity.
[Section 5.1] No event window is specified for the RGBE-Gaze experiments; if the same 33 ms window is used, state so explicitly, and otherwise provide the chosen window.
[Section 5.2] The RGBE-Gaze split is not described in terms of subjects; please provide the number of subjects in the training, validation, and test sets, and clarify the relationship between "target users" and evaluation subjects.
[Section 5.4] The sentence "The models were implemented in PyTorch and trained using distributed GPU execution for eBIRD-G and eBIRD-S training, validation and evaluation" is awkward; consider rephrasing to "distributed GPU execution for training, validation, and evaluation of eBIRD-G and eBIRD-S."
[References] References [12] and [13] appear to refer to the same E2VID publication with different titles; please consolidate them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; eBIRD is a standard DDPM/ControlNet application evaluated against external baselines.

full rationale

The paper contains no fitted parameter that is renamed as a prediction, no self-citation used as load-bearing evidence, and no uniqueness claim imported from the authors' prior work. The framework is assembled from published DDPM and ControlNet components (Ho et al. 2020; Zhang et al. 2023) with a conventional two-stage training objective (Eq. 8) and a standard conditional sampling procedure (Algorithm 2). The central quantitative claims (Tables 1 and 2) are empirical comparisons against external baselines E2VID, ET-Net, and Pix2Pix, so the reconstruction result is not equivalent by construction to the training objective. The conclusion that 'the preferred learning strategy depends on the reconstruction domain' is a post hoc reading of the same experimental numbers, but that is an interpretive summary, not a derivation that reduces to its own inputs. The only flagged weakness is the RGBE-Gaze protocol: Section 5.2 says eBIRD-S is 'further fine-tuned using data from the target subjects' without stating that fine-tuning subjects are disjoint from evaluation subjects; if they overlap, the specialized advantage would measure memorization rather than generalization. That is a validation and potential data-leakage concern, not a circularity in the mathematical derivation, and the paper itself defers ablations to future work. No circular step meeting the evidentiary bar was found.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical comparison between two training strategies; no physical constants or fitted parameters are introduced. The free parameters are experimental design choices. The main assumptions are imported from the cited DDPM and ControlNet literature, plus two unstated domain assumptions about the event-frame representation and the baseline protocols.

free parameters (2)

Event window T = 33 ms
Hand-chosen temporal window for all experiments; no comparison across window lengths is reported, so conclusions are conditional on this value.
Training hyperparameters = learning rate 1e-4, batch size 80, 40 epochs
Shared across eBIRD-G and eBIRD-S, so not load-bearing for the strategy comparison, but hand-chosen without sensitivity analysis.

assumptions (4)

standard math DDPM forward and reverse processes (Eqs. 3-6) correctly model the image distribution
Standard result from Ho et al. [7]; cited and used without proof.
domain assumption ControlNet zero-convolution conditioning (Eq. 7) can inject event frames without degrading the pretrained image prior
Imported from Zhang et al. [20]; central to the framework, not independently verified here.
domain assumption A single accumulated event frame over 33 ms is a sufficient representation of the event stream for reconstruction
Used for all experiments; the paper does not ablate this representation.
domain assumption Pretrained E2VID and ET-Net weights are directly applicable to RGBE-Gaze event frames without adaptation
The baselines are evaluated out of the box; this is likely mismatched and affects the comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of eBIRD: Event-based Intensity Image Reconstruction Using Controllable Diffusion Models." pith.science (2026). https://pith.science/paper/G5JVIOTY

@misc{pith2026260808519,
  author       = {Pith},
  title        = {Pith review of: eBIRD: Event-based Intensity Image Reconstruction Using Controllable Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G5JVIOTY}},
  note         = {Machine review of arXiv:2608.08519}
}

read the original abstract

Intensity-image reconstruction from event streams remains a challenging problem due to the binary, sparse, and asynchronous nature of event data. This work proposes eBIRD, an event-guided reconstruction framework that combines a DDPM with ControlNet-based conditioning. We analyze generic and specialized diffusion learning strategies for handwritten digit (N-MNIST) and face (RGBE-Gaze) reconstruction using 33ms event windows. On N-MNIST, the general model achieves the best reconstruction quality (MSE 0.0052, SSIM 0.8982, PSNR 23.34dB), whereas the specialized model performs best on RGBE-Gaze (MSE 0.0161, SSIM 0.7605, PSNR 19.08dB). These preliminary results suggest that controllable diffusion models are a promising approach for event-guided intensity-image reconstruction, while highlighting that the preferred learning strategy depends on the reconstruction domain.

Figures

Figures reproduced from arXiv: 2608.08519 by the authors.

**Figure 1.** Overview of the proposed eBIRD framework. A pretrained DDPM backbone is conditioned by a trainable ControlNet branch using event frames. Training consists of generic image-prior learning followed by event-guided specialization for the target reconstruction task. 4.2 Network Architecture Building Blocks. The proposed framework adopts the standard DDPM U-Net backbone together with a ControlNet branch of identical topo… view at source ↗

**Figure 2.** Detailed architecture of the DDPM backbone and ControlNet blocks. Each downsampling, bottleneck, and upsampling module comprises residual convolutional layers with group normalization, SiLU activation, self-attention, and timestep conditioning through sinusoidal positional embeddings. ior throughout the diffusion process. The SiLU activation function SiLU(x) = x σ(x) is used throughout the architecture. 4.3 Trainin… view at source ↗

**Figure 3.** Representative samples from the N-MNIST dataset [10]. Each pair consists of the original MNIST grayscale digit (left) and its corresponding event-based representation (right), illustrating the input-target pairs used for handwritten digit reconstruction. To compare the performance of class-specific and general reconstruction models, we designed 11 experimental scenarios: 10 models trained to reconstruct a single … view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Samples from the RGBE-Gaze dataset [21]. Each example shows the RGB target image with the corresponding event stream overlaid. For the experiments presented in this paper, RGB images are converted to grayscale and used as targets. 5.2 Training and Evaluation Details A…

**Figure 5.** Figure 5: Qualitative reconstruction results on the N-MNIST dataset using 33 ms event windows. The first column shows the event representation, followed by the reconstructions obtained with the class-specific models and the general model, together with the corresponding ground-…

**Figure 6.** Figure 6: Qualitative comparison of reconstruction methods on the RGBE-Gaze dataset. From left to right: input event representation, Pix2Pix reconstruction, eBIRD generic reconstruction (eBIRD-G), eBIRD specialized reconstruction (eBIRD-S), and the corresponding grayscale ground…

**Figure 7.** Figure 7: Qualitative reconstruction results on the RGBE-Gaze dataset. Each triplet shows the event representation (left), the image reconstructed by the proposed eBIRD framework (center), and the corresponding grayscale ground-truth image (right). Representative examples from …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 10 canonical work pages

[1]

Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2016).https://doi.org/10

Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2016).https://doi.org/10. 1109/cvpr.2016.102,https://doi.org/10.1109/cvpr.2016.1022 eBIRD 15

work page arXiv 2016
[2]

IEEE Transactions on Image Processing30, 2488–2500 (1 2021).https://doi.org/10.1109/tip.2021

Cadena,P.R.G.,Qian,Y.,Wang,C.,Yang,M.:SPADE-E2VID:Spatially-Adaptive Denormalization for Event-Based Video Reconstruction. IEEE Transactions on Image Processing30, 2488–2500 (1 2021).https://doi.org/10.1109/tip.2021. 3052070,https://doi.org/10.1109/tip.2021.30520702

arXiv 2021
[3]

Information15(8) (2024).https://doi.org/10.3390/ info15080472,https://www.mdpi.com/2078-2489/15/8/4721

Cazzato, D., Bono, F.: An application-driven survey on event-based neuromor- phic computer vision. Information15(8) (2024).https://doi.org/10.3390/ info15080472,https://www.mdpi.com/2078-2489/15/8/4721

work page 2024
[4]

In: European Conference on Computer Vision

Chakravarthi, B., Verma, A.A., Daniilidis, K., Fermuller, C., Yang, Y.: Recent event camera innovations: A survey. In: European Conference on Computer Vision. pp. 342–376. Springer (2024) 1

work page 2024
[5]

In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W

Dhariwal, P., Nichol, A.Q.: Diffusion models beat GANs on image synthesis. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neu- ral Information Processing Systems (2021),https://openreview.net/forum?id= AAWuCvzaVt2

work page 2021
[6]

IEEE Transactions on Pattern Analysis and Machine Intel- ligence44(1), 154–180 (2022).https://doi.org/10.1109/TPAMI.2020.3008413 1, 3

Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event- based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intel- ligence44(1), 154–180 (2022).https://doi.org/10.1109/TPAMI.2020.3008413 1, 3

arXiv 2022
[7]

Advances in neural information processing systems33, 6840–6851 (2020) 1, 2, 3, 4

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020) 1, 2, 3, 4

work page 2020
[8]

In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with con- ditional adversarial networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5967–5976 (2017).https://doi.org/10.1109/ CVPR.2017.6329, 12

work page arXiv 2017

Show all 23 references

[9]

arXiv preprint arXiv:2407.08231 (2024) 2

Liang, J., Yu, B., Yang, Y., Han, Y., Shi, B.: E2vidiff: Perceptual events-to-video reconstruction using diffusion priors. arXiv preprint arXiv:2407.08231 (2024) 2

2024 arXiv
[10]

Frontiers in neuroscience 9, 437 (2015) 8

Orchard, G., Jayawant, A., Cohen, G.K., Thakor, N.: Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience 9, 437 (2015) 8

2015
[11]

In: 2024 5th International Conference on Com- puter Vision, Image and Deep Learning (CVIDL)

Quan, Y., Zhang, Y.: Image reconstruction approaches based on fusion of event stream and image frame: A survey. In: 2024 5th International Conference on Com- puter Vision, Image and Deep Learning (CVIDL). pp. 232–238 (2024).https: //doi.org/10.1109/CVIDL62147.2024.106041421, 2

2024
[12]

2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2019).https://doi.org/10

Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-To-Video: Bringing Modern Computer Vision to Event Cameras. 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2019).https://doi.org/10. 1109/cvpr.2019.00398,https://doi.org/10.1109/cvpr.20...

2022
[13]

IEEE Transactions on Pattern Analysis and Machine Intelligence43(6), 1964–1980 (2021).https://doi.org/10.1109/TPAMI

Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High speed and high dynamic range video with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence43(6), 1964–1980 (2021).https://doi.org/10.1109/TPAMI. 2019.29633861, 9, 12

2021
[14]

In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10674–10685 (2022). https://doi.org/10.1109/CVPR52688.2022.010421, 2

2022
[15]

In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention – MICCAI
[16]

Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R.E., Scaramuzza, D.: Fast Image Reconstruction with an Event Camera. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 156–163 (2020) pp. 156–163 (3 2020).https://doi.org/10.1109/...

2020
[17]

IEEE Access12, 51275–51306 (2024).https://doi.org/10.1109/ACCESS.2024.33860321

Shariff, W., Dilmaghani, M.S., Kielty, P., Moustafa, M., Lemley, J., Corcoran, P.: Event cameras in automotive sensing: A review. IEEE Access12, 51275–51306 (2024).https://doi.org/10.1109/ACCESS.2024.33860321

2024
[18]

Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using trans- former.In:2021IEEE/CVFInternationalConferenceonComputerVision(ICCV). pp. 2543–2552 (2021).https://doi.org/10.1109/ICCV48922.2021.002561, 2, 9, 12

2021
[19]

In: 2025 IEEE International Conference on Unmanned Systems (ICUS)

Zhang, L., Yang, P., Yang, H., Wang, F., Zhang, S.: Review on event-based vision: Data representation, image generation, calibration and recognition. In: 2025 IEEE International Conference on Unmanned Systems (ICUS). pp. 1030–1040 (2025). https://doi.org/10.1109/ICUS66297.2025...

2025
[20]

In: 2023 IEEE/CVF International Conference on Computer Vi- sion (ICCV)

Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: 2023 IEEE/CVF International Conference on Computer Vi- sion (ICCV). pp. 3813–3824 (2023).https://doi.org/10.1109/ICCV51070.2023. 003551, 2, 4

2023
[21]

IEEE Transactions on Pattern Analysis and Machine Intelligence47(1), 601–615 (2025).https://doi.org/10.1109/TPAMI.2024.34748588, 9

Zhao, G., Shen, Y., Zhang, C., Shen, Z., Zhou, Y., Wen, H.: Rgbe-gaze: A large- scale event-based multimodal dataset for high frequency remote gaze tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence47(1), 601–615 (2025).https://doi.org/10.1109/TPAMI.2024....

2025
[22]

In: European Conference on Computer Vision

Zhu, L., Zheng, Y., Zhang, Y., Wang, X., Wang, L., Huang, H.: Temporal residual guided diffusion framework for event-driven video reconstruction. In: European Conference on Computer Vision. pp. 411–427. Springer (2024) 2

2024
[2015]

pp. 234–241. Springer International Publishing, Cham (2015) 4 16 I. Bugueno-Cordova et al

2015

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2016).https://doi.org/10

Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) (6 2016).https://doi.org/10. 1109/cvpr.2016.102,https://doi.org/10.1109/cvpr.2016.1022 eBIRD 15

work page arXiv 2016

[2] [2]

IEEE Transactions on Image Processing30, 2488–2500 (1 2021).https://doi.org/10.1109/tip.2021

Cadena,P.R.G.,Qian,Y.,Wang,C.,Yang,M.:SPADE-E2VID:Spatially-Adaptive Denormalization for Event-Based Video Reconstruction. IEEE Transactions on Image Processing30, 2488–2500 (1 2021).https://doi.org/10.1109/tip.2021. 3052070,https://doi.org/10.1109/tip.2021.30520702

arXiv 2021

[3] [3]

Information15(8) (2024).https://doi.org/10.3390/ info15080472,https://www.mdpi.com/2078-2489/15/8/4721

Cazzato, D., Bono, F.: An application-driven survey on event-based neuromor- phic computer vision. Information15(8) (2024).https://doi.org/10.3390/ info15080472,https://www.mdpi.com/2078-2489/15/8/4721

work page 2024

[4] [4]

In: European Conference on Computer Vision

Chakravarthi, B., Verma, A.A., Daniilidis, K., Fermuller, C., Yang, Y.: Recent event camera innovations: A survey. In: European Conference on Computer Vision. pp. 342–376. Springer (2024) 1

work page 2024

[5] [5]

In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W

Dhariwal, P., Nichol, A.Q.: Diffusion models beat GANs on image synthesis. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neu- ral Information Processing Systems (2021),https://openreview.net/forum?id= AAWuCvzaVt2

work page 2021

[6] [6]

IEEE Transactions on Pattern Analysis and Machine Intel- ligence44(1), 154–180 (2022).https://doi.org/10.1109/TPAMI.2020.3008413 1, 3

Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event- based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intel- ligence44(1), 154–180 (2022).https://doi.org/10.1109/TPAMI.2020.3008413 1, 3

arXiv 2022

[7] [7]

Advances in neural information processing systems33, 6840–6851 (2020) 1, 2, 3, 4

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020) 1, 2, 3, 4

work page 2020

[8] [8]

In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with con- ditional adversarial networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5967–5976 (2017).https://doi.org/10.1109/ CVPR.2017.6329, 12

work page arXiv 2017