Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

How Real is CARLAs Dynamic Vision Sensor? A Study on the Sim-to-Real Gap in Traffic Object Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Models trained solely on CARLA's synthetic DVS event data reach only 4.26 mAP on real traffic streams, while adding real data lifts performance to 15.69 mAP, quantifying a substantial sim-to-real gap.

desk verdict A useful CARLA DVS traffic dataset and training-mix sweep, but the headline sim-to-real claim is untested because the test sets are fixed and fully real, and the split accounting does not add up. read the letter →

arxiv 2506.13722 v1 pith:VG4EALOE submitted 2025-06-16 cs.CV

classification cs.CV
keywords eventcamerasdynamicvisionsensorCARLAsimulatorsim-to-realgapevent-basedobjectdetectiontrafficmonitoringrecurrenttransformersyntheticdata
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 sets out to measure how much real-world performance is lost when an event-camera object detector is trained only on synthetic data from CARLA's built-in dynamic vision sensor (DVS). Training a recurrent vision transformer on purely synthetic clips and testing it on real event streams yields a mean average precision of 4.26, versus 15.69 when 85.7% of the training data is real. The authors interpret the monotonic rise in performance as the first quantifiable sim-to-real gap for CARLA DVS in traffic object detection, and they conclude that synthetic CARLA events cannot yet substitute for real event camera data in this setting. The paper also contributes SeTraM, a synthetic event-based traffic monitoring dataset with day/night and multi-intersection coverage.

What carries the argument

The analysis is carried by seven training datasets built by time-aligning synthetic SeTraM clips (from CARLA's sensor.camera.dvs) with real eTram clips, so that only the real/synthetic mix changes while total duration stays around 2300 seconds. All models use the RVT-Small recurrent vision transformer, which ingests event streams in 1MPX-format .npy/.h5 files, and evaluation uses mAP and AP@50 on fixed validation and test sets. The time-based alignment is the operative design choice that lets the authors swap data sources while keeping duration fixed.

What would settle it

Measure event density, object-size distributions, and class balance in SeTraM versus eTram clips, and check the timestamp ranges of the training and test splits for overlap. If the synthetic clips turn out to have far fewer events per second, smaller objects, or shared sequence IDs with the test set, the gap in Table 4 would reflect data statistics or leakage rather than DVS fidelity.

Watch

Extended reading notes

Core claim

The central claim is that models trained exclusively on CARLA DVS synthetic data perform well when tested on synthetic-heavy data but degrade significantly as the test stream becomes more real, while models trained on real-world data generalize more strongly across domains. The evidence is a controlled sweep: seven training sets with identical total duration differ only in the proportion of real eTram data (0%, 14.3%, ..., 85.7%), with fixed validation and test sets. On the mixed real-plus-synthetic test set, mAP rises from 4.26 for the synthetic-only model to 15.69 for the most real-heavy model; the night-only test set shows the same direction, from 4.22 to 12.63. The authors present this as the first quantitative demonstration that CARLA's native DVS module, despite its controllability and photorealistic environment, leaves a substantial domain gap in traffic object detection.

Load-bearing premise

The paper's conclusion rests on the assumption that the synthetic and real clips are interchangeable except for sensor realism, meaning same viewpoint, classes, annotation quality, and difficulty, and that the training, validation, and test clips are disjoint.

Editorial extensions

If this is right

  • Synthetic-only CARLA DVS training is insufficient for real-world traffic object detection; the best fully synthetic model reaches only 4.26 mAP on the mixed test set.
  • Each increment of real training data yields measurable gains, so even small amounts of annotated real event data are valuable for closing the domain gap.
  • Models trained on real-dominant data generalize better across day/night test conditions, suggesting domain adaptation should focus on transferring real-world robustness into synthetic training.
  • SeTraM can serve as a controlled benchmark for future sim-to-real studies, but it should not be treated as a drop-in replacement for real event data.
  • The near-linear trend on real test data implies a quantitative target: simulator improvements must raise synthetic-only performance toward the 15.69 mAP achieved with 85.7% real data.

Reading between the lines

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

  • Because the paper aligns synthetic and real clips by time rather than by event count or object instances, the measured slope may partly reflect differences in event density and scene complexity, not purely sensor realism; re-running the sweep with event-count-matched subsets would test this.
  • If the gap is driven by low-level event statistics such as sparsity, noise, and temporal uniformity rather than semantic content, then post-processing synthetic streams with realistic noise models or event-rate matching could recover much of the lost performance without new real data.
  • The validation peak at intermediate real proportions suggests a practical recipe for domain adaptation: a minority of real data can anchor the target domain while synthetic data provides scale and diversity, pointing to a sweet-spot mix rather than a rule that more real data is always better.
  • The event quality score mentioned in the paper could be applied to SeTraM clips to predict transferability before training, turning the binary result that a gap exists into a per-clip diagnostic.
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

4 major / 6 minor

Summary. The paper introduces SeTraM, a synthetic event-based traffic-monitoring dataset generated with CARLA's DVS, and trains an RVT-small detector on seven training sets that mix SeTraM and real eTram data in increasing real-data proportions (0% to 85.7%). It reports validation and test mAP/AP@50 and argues that the monotonic improvement in test performance as more real training data is added quantifies a sim-to-real gap, concluding that CARLA's synthetic events are not yet a substitute for real events in traffic object detection. The abstract additionally claims that synthetic-only models perform well on synthetic-heavy test sets and degrade as the real proportion in the test stream increases.

Significance. The empirical question—whether CARLA's built-in DVS can produce event streams useful for real traffic-object detection—is practically important, and the paper's controlled comparison across seven training compositions is a useful start. If the results were fully supported, the paper would provide a valuable benchmark for domain adaptation in neuromorphic traffic perception. The main strengths are the construction of a synthetic dataset with fixed overhead intersections, the use of a strong event-based detector (RVT), and the attempt to isolate the effect of real-data proportion by holding training duration constant. However, the headline claim about synthetic-heavy test sets is not measured by the reported experiments, and there is a data-accounting inconsistency that threatens the validity of the quantitative conclusions.

major comments (4)
  1. [Abstract; Section 3.3; Table 4] The abstract's central claim—that models trained solely on synthetic data 'perform well on synthetic-heavy test sets but suffer significant performance degradation as the proportion of real-world data increases'—is not tested by the reported experiments. Section 3.3 states that the validation and test sets remain fixed across all runs, and Section 3.1 describes the two test variants as subsets of real eTram data; in no experiment is the test-set composition varied. Table 4 therefore shows a training-data effect: a model trained with more real data scores higher on a fixed real test set. This is expected and does not quantify the sim-to-real gap of CARLA's DVS. To support the abstract claim, the authors would need to fix the training data (for example, Dataset #1, synthetic-only) and evaluate that fixed model on test sets with increasing real-data fraction, holding total duration and annotation format constant.
  2. [Section 3.1; Section 3.3; Table 2] There is a data-accounting inconsistency that risks train/test leakage. Section 3.3 gives seven groups of 333 seconds each, implying a total SeTraM duration of 2331 seconds (~38.9 minutes), which matches the '38 minutes' stated in Section 3.1. However, Dataset #1 in Table 2 uses approximately 2300 seconds of SeTraM for training, while Section 3.3 additionally allocates 320 seconds of SeTraM to validation and 160 seconds to test. The sum of these portions is 2780 seconds, which exceeds the stated total by roughly 450 seconds. Either the total duration, the split durations, or the disjointness of the splits is misreported; if the training and validation/test SeTraM clips overlap, the reported results would be invalid.
  3. [Section 5; Tables 3–4; Figure 9] The headline quantitative summary, 'an average slope of 0.115 mAP per unit increase in real data proportion,' cannot be reproduced from the reported tables. From Table 4, the mixed-test mAP increases from 4.26 (Dataset #1) to 15.69 (Dataset #7) over a real-proportion change of 0.857, which corresponds to roughly 13.3 mAP per unit (or about 0.133 per percentage point), not 0.115. Table 3's validation mAP is non-monotonic, so the slope cannot refer to that curve either. The authors should state exactly which curve the slope is fitted to and report the fitted value; as written, the central quantitative claim is unsupported.
  4. [Section 3.2; Section 3.3; Tables 2–4] The comparison does not establish that the observed performance gap is attributable to CARLA's DVS realism rather than to other dataset differences. Section 3.3 aligns SeTraM and eTram by temporal duration only, and no statistics are provided for event rate, event density, object-size distribution, class balance, annotation quality, camera perspective, or task difficulty. Without such characteristics, the monotone improvement in Table 4 could reflect differences in scene content or annotation protocols. This is load-bearing for the paper's claim to be a quantifiable sim-to-real gap rather than a dataset-comparison artifact.
minor comments (6)
  1. [Abstract; Section 1; Section 4] The abstract and introduction say a model is trained 'exclusively on synthetic data,' but Section 4 trains seven models on mixed datasets; only Dataset #1 is synthetic-only. The wording should be clarified.
  2. [Figure 9; Section 4.2] Figure 9's caption reads 'Validation mAP across models...' while the text describes it as test performance on fully real test sets; the caption and the referenced table should be made consistent.
  3. [Section 3.1; Section 3.3] The paper refers inconsistently to 'five daytime and two nighttime sequences' versus 'seven groups' of four intersection instances each; the terminology should be unified because a group is 333 seconds, not 80 seconds.
  4. [Throughout] The naming 'eTram' versus 'eTraM' is inconsistent (for example, Table 4 header versus Section 3.1), and the text contains typos such as 'CARLAs' in the abstract and '1M PXformat' in Section 3.2.
  5. [Section 4.1; Tables 3–4] No error bars or multiple seeds are reported; several adjacent configurations in Table 4 differ by less than 1.5 mAP, so the claimed monotonic trend should be supported by repeated runs or an explicit statement that the results are single-seed.
  6. [Table 3] Dataset #7's validation AP@75 collapses to 4.81 while its mAP is 12.21; the authors should discuss whether this is a metric artifact or a real failure mode of the heavily real-data model.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the claimed sim-to-real gap is an empirical measurement, though the abstract's 'synthetic-heavy test set' claim is not supported by the fixed-test design.

full rationale

This is an empirical benchmark study, not a derivation chain. The monotone rise in Table 4 mAP from 4.26 to 15.69 is a measured training-data effect on fixed real eTram test sets, and the validation slope of 0.115 is a descriptive summary, not a parameter used to generate predictions. No equation defines a target quantity in terms of its own inputs, and no fitted value is renamed as a prediction. The eTram, SeTraM, and EQS references include same-group work, but eTram is an independently captured published dataset and EQS is only mentioned as a future metric, so none of these citations is load-bearing. The abstract's statement that models were 'tested on varying combinations of synthetic and real-world event streams' is contradicted by Section 3.3, where validation and test sets are fixed; this is a correctness/validity concern about the central claim, but it is not circularity under the specified patterns.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The central claim rests on domain comparability between synthetic and real data, disjoint splits, and single-run training stability. No fitted physical parameters are introduced by the paper; the training hyperparameters are standard model configurations rather than parameters fitted to the headline result.

assumptions (4)
  • domain assumption eTraM is a representative real-world event camera benchmark for traffic monitoring.
    The paper's central gap measurement treats eTraM as ground truth for real event data; if eTraM annotations or capture conditions are unrepresentative, the gap numbers do not generalize. See Section 3.
  • domain assumption Time-based alignment makes SeTraM and eTraM domain-comparable while controlling for event density.
    The paper aligns by duration, not by event count or object distribution, and reports no statistics on class frequency, object size, or event rate, so the 'gap' may include dataset-specific confounds. See Section 3.3.
  • domain assumption Training, validation, and test subsets are disjoint.
    The stated SeTraM total (38 minutes) is less than the sum of training (about 2300 seconds), validation (320 seconds), and test (160 seconds), so the paper implicitly assumes extra disjoint data or a reporting error; this premise is load-bearing. See Sections 3.1 and 3.3.
  • domain assumption Single-run RVT training is stable enough that mAP differences are meaningful.
    No multiple seeds or confidence intervals are reported; large jumps, such as AP@50 from 26.07 to 67.77, suggest possible instability. See Section 4.1 and Table 3.
invented entities (1)
  • SeTraM dataset
    purpose: Synthetic event-based traffic monitoring dataset used to train models and quantify the sim-to-real gap.
    No public URL or release mechanism is provided in the text; the paper describes generation scripts but does not ship them, so the dataset cannot be independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Real is CARLAs Dynamic Vision Sensor? A Study on the Sim-to-Real Gap in Traffic Object Detection." pith.science (2026). https://pith.science/paper/VG4EALOE

@misc{pith2026250613722,
  author       = {Pith},
  title        = {Pith review of: How Real is CARLAs Dynamic Vision Sensor? A Study on the Sim-to-Real Gap in Traffic Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VG4EALOE}},
  note         = {Machine review of arXiv:2506.13722}
}
read the original abstract

Event cameras are gaining traction in traffic monitoring applications due to their low latency, high temporal resolution, and energy efficiency, which makes them well-suited for real-time object detection at traffic intersections. However, the development of robust event-based detection models is hindered by the limited availability of annotated real-world datasets. To address this, several simulation tools have been developed to generate synthetic event data. Among these, the CARLA driving simulator includes a built-in dynamic vision sensor (DVS) module that emulates event camera output. Despite its potential, the sim-to-real gap for event-based object detection remains insufficiently studied. In this work, we present a systematic evaluation of this gap by training a recurrent vision transformer model exclusively on synthetic data generated using CARLAs DVS and testing it on varying combinations of synthetic and real-world event streams. Our experiments show that models trained solely on synthetic data perform well on synthetic-heavy test sets but suffer significant performance degradation as the proportion of real-world data increases. In contrast, models trained on real-world data demonstrate stronger generalization across domains. This study offers the first quantifiable analysis of the sim-to-real gap in event-based object detection using CARLAs DVS. Our findings highlight limitations in current DVS simulation fidelity and underscore the need for improved domain adaptation techniques in neuromorphic vision for traffic monitoring.

Figures

Figures reproduced from arXiv: 2506.13722 by the authors.

Figure 1
Figure 1. Examples of traffic objects captured using event-based vision. (Left, black background) Synthetic events generated by CARLA’s DVS module. (Right, gray back￾ground) Real-world events from the eTram dataset [23], recorded using a physical event camera (Prophesee’s EVK 4 HD). asynchronously record changes in pixel intensity, producing a continuous stream of events, making them well-suited for real-time object detection… view at source ↗
Figure 2
Figure 2. Off-the-shelf event cameras. (a) Prophesee family of event cameras [18], (left to right): Century Arks SilkyEVCam, OpenMV GENX320 Camera Module, IDS uEye XCP, and Lucid Triton2 EVS. (b) IniVation family of event cameras [11], (left to right): DVXplorer, DVXplorer Micro, DAVIS346, and DVXplorer Lite. This sensing paradigm offers several advantages over conventional RGB cam￾eras, including high dynamic range, negligib… view at source ↗
Figure 3
Figure 3. Event frames (top row) generated by CARLA’s DVS and their corresponding RGB frames (bottom row), shown across different viewpoints with traffic objects. positive and negative events are typically rendered in blue and red, respectively (See [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Snapshots of SeTram dataset showing daytime, nighttime and annotation in￾stances. SeTraM consists of five daytime and two nighttime sequences, each lasting approximately 80 seconds, for a total of 38 minutes of event data. To emulate realistic traffic dynamics, up to 1…
Figure 5
Figure 5. Figure 5: Data generation pipeline for the training sets from the CARLA simulator at four-way intersections and the real-world eTram dataset – Traffic Simulation: Using CARLA’s Python API, dynamic urban scenarios were constructed with moving vehicles and pedestrians. Four inters…
Figure 6
Figure 6. Figure 6: Validation mAP across models with increasing real-world data proportions. Precision (mAP). As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Validation loss curves across different training configurations. The [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Training loss vs. epochs for each dataset configuration [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Validation mAP across models with increasing real-world data proportions. Training loss steadily decreased across all runs, indicating effective opti￾mization, while validation loss fluctuated more substantially, likely due to do￾main shifts between synthetic and real …
Figure 10
Figure 10. Figure 10: Qualitative visualization of predictions for models 1 to 7. Ground-truth labels are shown in the bottom row, and model predictions are displayed in the top row. ability (e.g., lighting conditions) still poses a challenge for sim-to-real transfer [PITH_FULL_IMAGE:figu…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Event-RGB Adaptive Tracking for Nighttime Highway Perception

    cs.CV 2026-07 conditional novelty 5.5 of 10

    JEAT jointly associates RGB and event detections with NIS-adapted measurement noise, raising MOTA on unlit nighttime highways from 46% (RGB) / 69% (event) to 77% on a new CARLA dataset.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2404.10540 (2024)

    Aliminati, M.R., Chakravarthi, B., Verma, A.A., Vaghela, A., Wei, H., Zhou, X., Yang, Y.: Sevd: Synthetic event-based vision dataset for ego and fixed traffic per- ception. arXiv preprint arXiv:2404.10540 (2024)

  2. [2]

    In: 2017 IEEE International Conference on Image Processing (ICIP)

    Bi, Y., Andreopoulos, Y.: Pix2nvs: Parameterized conversion of pixel-domain video frames to neuromorphic vision streams. In: 2017 IEEE International Conference on Image Processing (ICIP). pp. 1990–1994. IEEE (2017)

  3. [3]

    In: International Conference on Interdisciplinary Approaches in Civil Engineering for Sustainable Development

    Chakravarthi, B., Manoj Kumar, M., Pavan Kumar, B.: Event-based sensing for improved traffic detection and tracking in intelligent transport systems toward sustainable mobility. In: International Conference on Interdisciplinary Approaches in Civil Engineering for Sustainable Development. pp. 83–95. Springer (2023)

  4. [4]

    arXiv preprint arXiv:2408.13627 (2024)

    Chakravarthi, B., Verma, A.A., Daniilidis, K., Fermuller, C., Yang, Y.: Recent event camera innovations: A survey. arXiv preprint arXiv:2408.13627 (2024)

  5. [5]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Chanda, K., Verma, A., Vaghela, A., Yang, Y., Chakravarthi, B.: Event quality score (eqs): Assessing the realism of simulated event camera streams via distance in latent space. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5105–5113 (2025)

  6. [6]

    In: Conference on robot learning

    Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: Carla: An open urban driving simulator. In: Conference on robot learning. pp. 1–16. PMLR (2017) This paper has been accepted at the 10th International Conference on Computer Vision & Image Processing (CVIP 2025) How Real is CARLA’s Dynamic Vision Sensor? 15

  7. [7]

    IEEE transactions on pattern analysis and machine intelligence 44(1), 154–180 (2020)

    Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154–180 (2020)

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Gehrig, M., Scaramuzza, D.: Recurrent vision transformers for object detection with event cameras. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13884–13893 (2023)

Show all 25 references
  1. [9]

    In: European Conference on Computer Vision

    Han, H., Lyu, J., Li, J., Wei, H., Li, C., Wei, Y., Chen, S., Ji, X.: Physical-based event camera simulator. In: European Conference on Computer Vision. pp. 19–35. Springer (2024)

  2. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 1312–1321 (2021)

  3. [11]

    https://inivation.com/buy/, accessed: 2025-04-16

    iniVation AG: Buy – inivation. https://inivation.com/buy/, accessed: 2025-04-16

  4. [12]

    https://inivation.com/solutions/, accessed: 2025-04-16

    iniVation AG: Solutions – inivation. https://inivation.com/solutions/, accessed: 2025-04-16

  5. [13]

    arXiv preprint arXiv:2411.12250 (2024)

    Jiang, X., Zhou, F., Lin, J.: Adv2e: Bridging the gap between analogue circuit and discrete frames in the video-to-events simulator. arXiv preprint arXiv:2411.12250 (2024)

  6. [14]

    Frontiers in Neuroscience 15, 702765 (2021)

    Joubert, D., Marcireau, A., Ralph, N., Jolley, A., Van Schaik, A., Cohen, G.: Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience 15, 702765 (2021)

  7. [15]

    In: European Conference on Computer Vi- sion

    Lin, S., Ma, Y., Guo, Z., Wen, B.: Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors. In: European Conference on Computer Vi- sion. pp. 578–593. Springer (2022)

  8. [16]

    The International journal of robotics research36(2), 142–149 (2017)

    Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event- camera dataset and simulator: Event-based data for pose estimation, visual odome- try, and slam. The International journal of robotics research36(2), 142–149 (2017)

  9. [17]

    Advances in Neural Information Processing Systems 33, 16639–16652 (2020)

    Perot, E., De Tournemire, P., Nitti, D., Masci, J., Sironi, A.: Learning to detect ob- jects with a 1 megapixel event camera. Advances in Neural Information Processing Systems 33, 16639–16652 (2020)

  10. [18]

    https://www.prophesee.ai/event- based-evaluation-kits/, accessed: 2025-04-16

    Prophesee: Event-based vision evaluation kits. https://www.prophesee.ai/event- based-evaluation-kits/, accessed: 2025-04-16

  11. [19]

    https://www.prophesee.ai/event-based-vision-more-applications/ (2024), ac- cessed: 2024-04-14

    Prophesee: Event-based vision applications | prophesee. https://www.prophesee.ai/event-based-vision-more-applications/ (2024), ac- cessed: 2024-04-14

  12. [20]

    https://docs.prophesee.ai/stable/ appli- cations.html (2024), accessed: 2025-04-16

    Prophesee: Video to event gpu simulator. https://docs.prophesee.ai/stable/ appli- cations.html (2024), accessed: 2025-04-16

  13. [21]

    In: Conference on robot learning

    Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969–982. PMLR (2018)

  14. [22]

    IEEE Access (2024)

    Shariff, W., Dilmaghani, M.S., Kielty, P., Moustafa, M., Lemley, J., Corcoran, P.: Event cameras in automotive sensing: A review. IEEE Access (2024)

  15. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Verma, A.A., Chakravarthi, B., Vaghela, A., Wei, H., Yang, Y.: etram: Event- based traffic monitoring dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22637–22646 (2024)

  16. [24]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub, U., Rahman, T.: V2ce: Video to continuous events simulator. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 12455–12461. IEEE (2024)

  17. [25]

    In: 2021 IEEE international conference on com- putational photography (ICCP)

    Zhu, A.Z., Wang, Z., Khant, K., Daniilidis, K.: Eventgan: Leveraging large scale image datasets for event cameras. In: 2021 IEEE international conference on com- putational photography (ICCP). pp. 1–11. IEEE (2021)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.