Pith. sign in

REVIEW 3 major objections 4 minor 78 references

E2Pano reconstructs full 360-degree panoramas from rotational event scans using a geometry-guided learned pipeline, reporting large quality gains over optimization-based baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

E2Pano couples spherical event alignment with a learned Transformer-based photometric stage to reconstruct panoramas from event streams under rotational scanning, outperforming optimization baselines on a new synthetic benchmark.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The pipeline design and dataset are genuinely useful, but the headline 13 dB win in Table 1 may be an artifact of unregistered baseline outputs and needs verification before it can be believed. the 3 major comments →

arxiv 2608.00694 v1 pith:WJUG7LUI submitted 2026-08-01 cs.CV

E2Pano: Learning Event-to-Panorama Image Reconstruction

Zhenyang Li , Zongqi He , Jia Pan , Shijie Lin , Yifan Peng This is my paper
classification cs.CV
keywords event cameraspanoramic reconstructionspherical mappingevent-to-imagerotational scanningtransformerfrequency-domain lossdataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Event cameras capture brightness changes at microsecond resolution and high dynamic range, which should make motion-blur-free panoramic imaging possible by simply rotating the camera. The obstacle is reconstruction: optimization-based methods that jointly estimate motion and photometry are slow, and learning-based event-to-image models assume flat perspective images. E2Pano claims to remove that bottleneck by first aligning events on a sphere with an existing rotation estimator and then learning to reconstruct the panorama from spherical event maps, keeping the true 3D viewing directions throughout. On its synthetic benchmark it reports large gains over the optimization and stitching baselines, and it transfers to real rotational captures despite being trained only on synthetic data. If correct, the method turns a long iterative optimization into a fast feed-forward reconstruction, bringing event-based panoramic capture closer to practical use.

Core claim

E2Pano's central claim is that a full 360°×180° luma panorama can be recovered from a rotational event stream by decoupling geometry from photometry: use spherical ICP to place every event on the unit sphere, then feed the accumulated sphere into an end-to-end learned reconstruction network rather than running photometric bundle adjustment over the raw events. The network consists of a roughly 15K-parameter residual enhancement module that turns sparse event accumulations into smoother, feature-rich inputs, trained with a frequency-domain loss that matches magnitude and phase spectra, and a spherical Transformer decoder whose positional embeddings come from the preserved 3D unit vectors, not

What carries the argument

The load-bearing object is the spherical geometry channel V(u,v): the exact 3D unit vector carried from the ES-ICP alignment stage into every later network layer. It is used as the positional embedding in the spherical Transformer, so the decoder always reasons about true viewing directions instead of distorted equirectangular coordinates. The other two components are the residual event enhancement module (with learnable weight α and frequency-domain supervision) and a frozen self-supervised visual encoder (DINOv2) that provides semantic features; together they bridge the 'event-image domain gap'—sparse, stepwise event accumulations versus continuous natural luminance.

Load-bearing premise

The whole method assumes the rotation-estimation step (spherical ICP) gives accurate camera rotations during scanning; if that alignment is wrong, the photometric reconstruction inherits the error and fails, as the paper acknowledges.

What would settle it

On a motorized turntable with ground-truth rotation angles, sweep angular velocity from 5 to 60 deg/s across scenes with varying texture density; if E2Pano's PSNR falls below EPBA's within the 10–30 deg/s range the paper claims, or if quality collapses at speeds where the geometric front-end still has overlapping observations, the central quality claim is falsified. A complementary check is a fully end-to-end timing run in which all baselines execute both alignment and photometric reconstruction on identical event streams; if E2Pano's speed advantage disappears, the efficiency claim is falsifi

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Rotational event scans can be turned into panoramas without iterative photometric optimization over the raw event stream, reducing the reconstruction bottleneck to a feed-forward pass over a spherical event map.
  • Training purely on synthetic events transfers to real captures from a different rig under the same camera-center-fixed rotational protocol, suggesting synthetic data can substitute for expensive real paired data.
  • The frequency-domain loss plus enhancement module recovers high-frequency detail and smooth gradients that raw event accumulations lack, which is the main source of the reported quality gain.
  • The pipeline's failure modes are inherited from the geometric front-end: high-speed rotation, repetitive or textureless regions, and multi-row scans that break the pure-rotation assumption degrade or break reconstruction.
  • PanoScan provides 4,370 synthetic and 30 real rotational event-panorama pairs as a benchmark for future event-to-panorama methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the learned photometric stage is agnostic to the source of the rotation estimate, the same back-end could be reused with a more robust 6-DoF geometric front-end; the paper's limitation to pure rotation is a property of the front-end, not the reconstruction network.
  • The reported speed comparison likely understates the advantage in realistic use, since the baseline timings cover only event alignment while E2Pano's 0.2 s includes the full pipeline; a fully standardized end-to-end benchmark on identical hardware would clarify the margin.
  • A natural extension is to condition the enhancement module on event rate or rotation speed, since the paper's failure cases concentrate where event support is sparse or overlap is low; this could be tested without changing the spherical architecture.
  • If the synthetic-to-real transfer holds across more diverse rigs and trajectories, event-based panorama capture could become a low-cost alternative to bracket HDR stitching for VR and telepresence content.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes E2Pano, a two-stage pipeline for reconstructing 360x180-degree luma panoramas from event streams acquired under rotational scanning. The geometric front-end uses the existing ES-ICP method to estimate SO(3) poses and accumulate events on a sphere; the photometric back-end is a learnable module that enhances the accumulated event map and reconstructs the panorama with a spherical Transformer using 3D positional embeddings, trained with L1, edge, and frequency-domain losses. The authors introduce the PanoScan dataset (4,370 synthetic scenes from OmniCity and 30 real captures) and report a large quantitative margin over optimization-based and stitching-based baselines on the synthetic test set (about 13 dB PSNR, 0.55 SSIM, 0.018 LPIPS), plus qualitative results on real captures.

Significance. If the reported gains are real, E2Pano would be a significant step toward practical event-based panoramic imaging, replacing expensive per-scene optimization with a learned reconstruction stage while retaining geometric awareness. The dataset and code release plans are also valuable to the community. The paper is honest about its operating boundary (pure rotation, dependence on ES-ICP) and includes machine-checked experiments, ablations, and failure cases. However, the central quantitative claim depends on a fair evaluation protocol, and the current manuscript does not make that protocol sufficiently explicit.

major comments (3)
  1. [Sec. 4.1 / Sec. 4.2, Table 1] The manuscript does not state that baseline outputs are registered to the ground-truth equirectangular coordinate frame before computing PSNR/SSIM/LPIPS. Each baseline estimates its own trajectory; for rotational SLAM, the global yaw offset is arbitrary, and a yaw shift of even a few degrees would collapse PSNR/SSIM in a 360-degree panorama. The near-zero SSIM for E2VID+Stitch (0.007) is consistent with a large offset rather than pure photometric failure. Please specify the alignment procedure (e.g., optimal yaw/pitch search, use of ground-truth poses for baselines, or cropping invalid border regions) and re-report the metrics. Without this, the 13 dB margin cannot be attributed to reconstruction quality.
  2. [Sec. 4.1, Baselines paragraph] The sentence 'We use ES-ICP for geometric initialization and evaluate the performance of the complete pipeline in all comparisons' is ambiguous. If ES-ICP is used to initialize the baselines, that would give the proposed pipeline an unfair advantage because ES-ICP is a prior method from the same group. If it is used only for E2Pano, the sentence should say so. Please clarify which methods receive ES-ICP initialization and whether the baselines run in their default configuration.
  3. [Sec. 4.6, Table 3] The timing comparison is not end-to-end: for CMax-SLAM and EPBA, only the event-alignment stage is timed, while for E2Pano the full pipeline (mapping + enhancement + reconstruction) is timed. The speedup numbers in Table 3 and the abstract's claim of 'lower photometric reconstruction cost' are therefore not directly supported. The paper acknowledges this, but the table's title and the '28.0x' speedup column are misleading. Please present photometric-stage-only timings separately, or report full end-to-end timings for all methods, and temper the speedup claim accordingly.
minor comments (4)
  1. [Sec. 3.2, Eq. (3)] The magnitude term of the frequency loss is written over the full spectrum, while the phase term is restricted to the frequency mask Omega. Please state explicitly that the magnitude loss also uses the same mask or define otherwise, to avoid inconsistency.
  2. [Sec. 4.1 and Fig. 7] The paper interchangeably uses 'E2VID' and 'E2VID++' and mentions a 'retrained threshold-robust E2VID setting'. Please clarify which exact model variant is used, whether it is retrained by the authors or the original checkpoint, and what 'threshold-robust' means for reproducibility.
  3. [Sec. 4.3 / Fig. 7, Fig. 9] The qualitative comparisons on ECRot and EROAM are presented without a clear statement of how the panoramas are co-registered (if at all). Since these datasets do not have pixel-aligned GT, please at least describe the reference frame used for visualization so that the reader can judge whether the comparison is fair.
  4. [Sec. 4.5, Table 2] The ablation results are reported on a validation set (10% of PanoScan), while the main results in Table 1 are on 50 test scenes. Please state the overlap (if any) between the validation set and the test set, and consider reporting ablations on the same test split for consistency.

Circularity Check

0 steps flagged

No significant circularity: photometric reconstruction is trained end-to-end against ground truth; the only self-citation (ES-ICP) is a non-load-bearing geometric front-end.

full rationale

The derivation chain is: raw events -> spherical projection with ES-ICP rotations (Eq. 1) -> enhancement (Eq. 2) -> frequency and spatial losses (Eqs. 3 and 7) -> spherical Transformer decoder (Eqs. 5-6) -> predicted luma panorama. Training is supervised by ground-truth panoramas with L1, edge, and frequency losses on held-out synthetic scenes; test PSNR/SSIM/LPIPS are external to the training objective. No equation defines a predicted quantity in terms of a fitted parameter or vice versa; the network weights are optimized and then evaluated on unseen data. The only self-citation is ES-ICP (Xing et al., 2026), prior work by overlapping authors used as the geometric front-end. It is not load-bearing for the paper's novel photometric claim: the paper explicitly states 'if geometric alignment fails, the subsequent photometric reconstruction inherits the error and may fail accordingly' (Sec. 3.1), acknowledging ES-ICP as an external dependency rather than deriving its success from it. No uniqueness theorem or ansatz is imported from the authors' prior work. The evaluation-protocol concern about whether baseline outputs were registered to the ground-truth frame before computing Table 1 is a validity/fairness risk, not definitional circularity: the text does not define the reported 13 dB gain in terms of the method's own outputs. Score 2 reflects the single non-load-bearing self-citation; no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The pipeline introduces no new physical entities. It relies on a fixed-optical-center pure-rotation assumption, on ES-ICP as a trusted geometric oracle, and on the representativeness of v2e-simulated events and OmniCity panoramas for real scenes. These domain assumptions are the main unmotivated inputs.

free parameters (4)
  • loss weights = lambda_L1=1.0, lambda_edge=0.5, lambda_freq=0.3
    Hand-chosen in Eq. 7; no sensitivity analysis is reported.
  • frequency mask threshold tau = 0.1
    Hand-chosen in Eq. 3 to filter low-magnitude frequencies.
  • enhancement scale alpha = initialized 0.1, converges to 0.3-0.5
    Learned parameter but its range is reported; effectively controls enhancement strength.
  • optimizer hyperparameters = batch 8, peak lr 5e-5, weight decay 1e-2
    Standard choices reported in Sec 4.1 without justification.
axioms (6)
  • domain assumption Camera motion is pure rotation around a fixed optical center during capture (Sec 3.1).
    All spherical mapping and ES-ICP estimation rely on this; the paper acknowledges multi-row scans that deviate from it break the pipeline.
  • domain assumption ES-ICP (Xing et al. 2026) yields accurate SO(3) rotation estimates for the event streams.
    The pipeline uses ES-ICP directly as geometric front-end, and the paper states alignment failure propagates into the photometric stage.
  • domain assumption v2e-simulated events on OmniCity panoramas are representative of real event streams for sim-to-real transfer.
    Training is purely synthetic; real-capture results are only qualitative and the paper flags the limited trajectory family.
  • domain assumption Frozen DINOv2 features encode semantics useful for photometric inference from sparse event accumulations.
    LSR relies on DINOv2 as a frozen encoder (Eq. 4) with no fine-tuning or ablation of this choice.
  • domain assumption Accumulating events on a unit sphere and converting to an equirectangular grid preserves the photometric information needed for reconstruction.
    The method discards raw event timing after mapping; the design assumes aggregated maps suffice for luma recovery.
  • standard math Standard FFT properties and the L1 norm behave as usual for the frequency-domain loss (Eq. 3).
    Routine mathematical background for the magnitude and phase losses.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of E2Pano: Learning Event-to-Panorama Image Reconstruction." pith.science (2026). https://pith.science/paper/WJUG7LUI

@misc{pith2026260800694,
  author       = {Pith},
  title        = {Pith review of: E2Pano: Learning Event-to-Panorama Image Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJUG7LUI}},
  note         = {Machine review of arXiv:2608.00694}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Event cameras offer microsecond-level temporal resolution and high dynamic range, potentially facilitating motion-blur-free panoramic imaging from fast rotational scanning. Nonetheless, existing optimization-based methods remain computationally demanding, while prior learning-based reconstruction methods are largely designed for perspective imagery and lack geometry-aware support for panoramic outputs. We present E2Pano, a geometry-guided event-to-panorama pipeline with an end-to-end learnable photometric reconstruction stage. Our framework preserves real spherical coordinates from geometric mapping throughout the pipeline, employs a lightweight enhancement module with frequency-domain supervision to bridge the event-image domain gap, and leverages a spherical Transformer with 3D positional embeddings for photometric reconstruction. Experiments on synthetic data and captured rotational scans show improved reconstruction quality and lower photometric reconstruction cost than optimization-based baselines, together with encouraging transfer to real captures under our acquisition protocol despite training purely on synthetic data. Additionally, we construct PanoScan, a dataset with 4,370 synthetic and 30 real-world panoramic scenes paired with event streams. Our dataset and code will be released.

Figures

Figures reproduced from arXiv: 2608.00694 by Jia Pan, Shijie Lin, Yifan Peng, Zhenyang Li, Zongqi He.

Figure 1
Figure 1. Figure 1: Event-to-Panorama (E2Pano) overview. Top: Our pipeline uses spherical projection to transform raw event data into geometrically consistent event point clouds, which are then recon￾structed into photometric panoramic images (8,631×2,000 resolution) with an end-to-end learnable spherical reconstruction stage (LSR). Bottom: Compared with optimization-based methods such as CMax-SLAM (Guo & Gallego, 2024)+SMT (… view at source ↗
Figure 2
Figure 2. Figure 2: E2Pano Framework Overview. Our two-stage pipeline transforms raw event streams into high-fidelity panoramic images through three stages: (1) Spherical Event Mapping projects events onto a unit sphere via ES-ICP (Xing et al., 2026), producing an equirectangular representa￾tion and preserved 3D coordinates; (2) Event Enhancement employs a lightweight residual network trained with frequency-domain loss to bri… view at source ↗
Figure 3
Figure 3. Figure 3: Event-Image Domain Gap. Comparison of raw (top) v.s. enhanced (bottom) events. Columns: (left) accumulated images with marked scanline and zoom-ins to red squares, (center) pixel intensity profiles showing stepwise quantization (blue, 16 levels) v.s. smooth gradients (or￾ange), (right) histograms demonstrating sparse clustering v.s. continuous distribution. Our enhance￾ment bridges the domain gap from spar… view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of the Event Enhancement Module. A lightweight residual network (∼15K params.) with the learnable weight α transforms sparse event intensities into feature-rich representations compatible with pretrained encoders. To bridge this domain gap, we introduce a lightweight neural module fθ that transforms sparse event representations into feature-rich inputs compatible with pretrained encoders while… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results on PanoScan Synthetic Test Set. For each scene, left to right: Ground truth panorama (Col. 1), E2Pano (Ours, Col. 2), E2VID (Rebecq et al., 2019) (E2VID++) + Stitch (Col. 3), CMax-SLAM (Guo & Gallego, 2024) + SMT (Kim et al., 2014) (Col. 4), and EPBA (Guo & Gallego, 2025) (Col. 5). Our method yields sharper and more globally consistent reconstructions across the 360◦ panoramic FoV. Even… view at source ↗
Figure 6
Figure 6. Figure 6: Real-World Acquisition Setup and Representative Captured-Scene Results. Left￾most: Our event-based panoramic acquisition system comprising a DVXplorer camera and a DJI RS 4 stabilizer, enabling practical scanning acquisition in the wild. Right: Representative quali￾tative results on a subset of real captured scenes. For each scene, the compared full pipelines are arranged left-to-right as Event image, E2Pa… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Comparison on the Public EROAM-Captured Benchmark. For each scene, the compared full pipelines are arranged left-to-right as E2Pano (Ours), E2VID (Rebecq et al., 2019) (retrained, E2VID++-style)+Stitch, EPBA (Guo & Gallego, 2025), and CMax-SLAM (Guo & Gallego, 2024)+SMT (Kim et al., 2014). For readability, each panorama visualization is cropped from a larger reconstruction. Gray regions in our … view at source ↗
Figure 8
Figure 8. Figure 8: Ablation Study. Top row, left to right: Baseline, Baseline + Ledge, Baseline + Lfreq, Baseline + Enhancement Module. Bottom row: Full Model that incorporates all components and the ground truth reference. Red squares denote the zoom-in region. PSNR (dB) scores are shown in corners. scene representation. Therefore, the final outputs reflect complete-system performance rather than a shared-front-end photomet… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative Comparison on the ECRot Dataset. Columns from left to right show the event image, E2Pano (Ours), EPBA (Guo & Gallego, 2025), and CMax-SLAM (Guo & Gallego, 2024). This comparison evaluates visual transfer beyond PanoScan-Sync on public real-captured event data under an adapted rotational-panorama protocol. For each reconstructed panorama, four zoom-in patches (marked with blue, dark red, orange,… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

78 extracted references · 71 canonical work pages

  1. [1]

    IEEE Journal of Solid-State Circuits , volume=

    A 128 128 120 dB 15 s Latency Asynchronous Temporal Contrast Vision Sensor , author=. IEEE Journal of Solid-State Circuits , volume=. 2008 , doi=

    work page 2008

  2. [2]

    IEEE Journal of Solid-State Circuits , volume=

    A 240 180 130 dB 3 s Latency Global Shutter Spatiotemporal Vision Sensor , author=. IEEE Journal of Solid-State Circuits , volume=. 2014 , doi=

    work page 2014

  3. [3]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Event-Based Vision: A Survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2022 , doi=

    work page 2022

  4. [4]

    IEEE Robotics and Automation Letters , volume=

    EVO: A Geometric Approach to Event-Based 6-DOF Parallel Tracking and Mapping in Real Time , author=. IEEE Robotics and Automation Letters , volume=. 2017 , doi=

    work page 2017

  5. [5]

    IEEE Robotics and Automation Letters , volume=

    Ultimate SLAM? Combining Events, Images, and IMU for Robust Visual SLAM in HDR and High-Speed Scenarios , author=. IEEE Robotics and Automation Letters , volume=. 2018 , doi=

    work page 2018

  6. [6]

    IEEE Transactions on Robotics , year=

    EKLT-VIO: Leveraging events and keypoint-level sparsity for robust visual-inertial odometry , author=. IEEE Transactions on Robotics , year=

    work page

  7. [7]

    IEEE Transactions on Robotics , volume=

    EROAM: Event-Based Camera Rotational Odometry and Mapping in Real Time , author=. IEEE Transactions on Robotics , volume=. 2026 , doi=

    work page 2026

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Events-To-Video: Bringing Modern Computer Vision to Event Cameras , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2019 , doi=

    work page 2019

  9. [9]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Fast Image Reconstruction with an Event Camera , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2020 , doi=

    work page 2020

  10. [10]

    International Journal of Robotics Research , volume=

    The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM , author=. International Journal of Robotics Research , volume=

    work page

  11. [11]

    IEEE Robotics and Automation Letters , volume=

    DSEC: A stereo event camera dataset for driving scenarios , author=. IEEE Robotics and Automation Letters , volume=

    work page

  12. [12]

    IEEE Robotics and Automation Letters , volume=

    The multivehicle stereo event camera dataset: An event camera dataset for 3D perception , author=. IEEE Robotics and Automation Letters , volume=. 2018 , publisher=

    2018

  13. [13]

    Foundations and Trends in Computer Graphics and Vision , volume=

    Image alignment and stitching: A tutorial , author=. Foundations and Trends in Computer Graphics and Vision , volume=

    work page

  14. [14]

    International Journal of Computer Vision , volume=

    Automatic Panoramic Image Stitching using Invariant Features , author=. International Journal of Computer Vision , volume=. 2007 , doi=

    work page 2007

  15. [15]

    ECCV , pages=

    Omnidepth: Dense depth estimation for indoors spherical panoramas , author=. ECCV , pages=

    work page

  16. [16]

    CVPR , pages=

    360MonoDepth: High-resolution 360 ^ monocular depth estimation , author=. CVPR , pages=

    work page

  17. [17]

    CVPR , pages=

    BiFuse: Monocular 360 ^ depth estimation via bi-projection fusion , author=. CVPR , pages=

    work page

  18. [18]

    arXiv preprint arXiv:2509.26618 , year=

    DA ^2 : Depth Anything in Any Direction , author=. arXiv preprint arXiv:2509.26618 , year=

    arXiv

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    OmniCity: Omnipotent City Understanding With Multi-Level and Multi-View Images , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2023 , doi=

    work page 2023

  20. [20]

    SIGGRAPH , pages=

    Recovering high dynamic range radiance maps from photographs , author=. SIGGRAPH , pages=

    work page

  21. [21]

    ACM Transactions on Graphics , volume=

    Jump: Virtual Reality Video , author=. ACM Transactions on Graphics , volume=. 2016 , doi=

    work page 2016

  22. [22]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Semantic Visual Localization , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2018 , doi=

    work page 2018

  23. [23]

    IEEE Robotics and Automation , year=

    Telepresence systems with 360 ^ panoramic imaging , author=. IEEE Robotics and Automation , year=

    work page

  24. [24]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Event-based vision: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    work page

  25. [25]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Event-based visual flow , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=

    work page

  26. [26]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2018 , doi=

    work page 2018

  27. [27]

    IROS , pages=

    Fast event-based Harris corner detection exploiting the advantages of event-driven cameras , author=. IROS , pages=

    work page

  28. [28]

    CVPR , pages=

    HATS: Histograms of averaged time surfaces for robust event-based object classification , author=. CVPR , pages=

    work page

  29. [29]

    Frontiers in Neuroscience , volume=

    Converting static image datasets to spiking neuromorphic datasets using saccades , author=. Frontiers in Neuroscience , volume=

    work page

  30. [30]

    CVPR Workshops , pages=

    EV-SegNet: Semantic segmentation for event-based cameras , author=. CVPR Workshops , pages=

    work page

  31. [31]

    ECCV , pages=

    ESS: Learning event-based semantic segmentation from still images , author=. ECCV , pages=

    work page

  32. [32]

    International Journal of Computer Vision , volume=

    EMVS: Event-based multi-view stereo---3D reconstruction with an event camera in real-time , author=. International Journal of Computer Vision , volume=

    work page

  33. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Simultaneous 3D reconstruction and panorama generation using a rotating DVS camera , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    work page

  34. [34]

    IEEE Transactions on Robotics , volume=

    Event-Based Stereo Visual Odometry , author=. IEEE Transactions on Robotics , volume=. 2021 , doi=

    work page 2021

  35. [35]

    British Machine Vision Conference , year=

    Simultaneous mosaicing and tracking with an event camera , author=. British Machine Vision Conference , year=

    work page

  36. [36]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Simultaneous Optical Flow and Intensity Estimation from an Event Camera , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2016 , doi=

    work page 2016

  37. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Event-Based High Dynamic Range Image and Very High Frame Rate Video Generation Using Conditional Generative Adversarial Networks , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2019 , doi=

    work page 2019

  38. [38]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages=

    Reducing the Sim-to-Real Gap for Event Cameras , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages=. 2020 , doi=

    work page 2020

  39. [39]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Event-Based Video Reconstruction Using Transformer , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=. 2021 , doi=

    work page 2021

  40. [40]

    Proceedings of the 31st ACM International Conference on Multimedia , pages=

    Event-Diffusion: Event-Based Image Reconstruction and Restoration with Diffusion Models , author=. Proceedings of the 31st ACM International Conference on Multimedia , pages=. 2023 , doi=

    work page 2023

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    EventNeRF: Neural Radiance Fields From a Single Colour Event Camera , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2023 , doi=

    work page 2023

  42. [42]

    International Journal of Computer Vision , volume=

    Systems and Experiment Paper: Construction of Panoramic Image Mosaics with Global and Local Alignment , author=. International Journal of Computer Vision , volume=. 2000 , doi=

    work page 2000

  43. [43]

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

    Seamless Image Stitching of Scenes with Large Motions and Exposure Differences , author=. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=. 2006 , doi=

    work page 2006

  44. [44]

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

    Eliminating Ghosting and Exposure Artifacts in Image Mosaics , author=. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=. 2001 , doi=

    work page 2001

  45. [45]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Deep Rectangling for Image Stitching: A Learning Baseline , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2022 , doi=

    work page 2022

  46. [46]

    ACM Transactions on Graphics , volume=

    Content-aware rotation , author=. ACM Transactions on Graphics , volume=

    work page

  47. [47]

    ECCV , pages=

    Acquisition of localization confidence for accurate object detection , author=. ECCV , pages=

  48. [48]

    International Journal of Computer Vision , volume=

    A theory of single-viewpoint catadioptric image formation , author=. International Journal of Computer Vision , volume=

    work page

  49. [49]

    CVPR , pages=

    OmniFusion: 360 monocular depth estimation via geometry-aware fusion , author=. CVPR , pages=

    work page

  50. [50]

    ECCV , year=

    BiFuseV2: 360 ^ depth and normal estimation from a single RGB image , author=. ECCV , year=

    work page

  51. [51]

    ICML Workshops , year=

    DDD17: End-to-end DAVIS driving dataset , author=. ICML Workshops , year=

    work page

  52. [52]

    Frontiers in Neuroscience , volume=

    DVS benchmark datasets for object tracking, action recognition, and object recognition , author=. Frontiers in Neuroscience , volume=

    work page

  53. [53]

    CVPR Workshops , year=

    N-EPIC: A large-scale event-based dataset for egocentric perception , author=. CVPR Workshops , year=

    work page

  54. [54]

    Pattern Recognition , volume=

    Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks , author=. Pattern Recognition , volume=

    work page

  55. [55]

    arXiv preprint arXiv:2205.03467 , year=

    EVIMO2: An event camera dataset for motion segmentation, optical flow, structure from motion, and visual inertial odometry in indoor scenes with monocular or stereo algorithms , author=. arXiv preprint arXiv:2205.03467 , year=

    Pith/arXiv arXiv

  56. [56]

    Transactions on Machine Learning Research , year=

    DINOv2: Learning Robust Visual Features without Supervision , author=. Transactions on Machine Learning Research , year=

  57. [57]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages=

    v2e: From Video Frames to Realistic DVS Events , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages=. 2021 , doi=

    work page 2021

  58. [58]

    High Speed and High Dynamic Range Video with an Event Camera , journal =

    Henri Rebecq and Ren. High Speed and High Dynamic Range Video with an Event Camera , journal =

    work page

  59. [59]

    IEEE Transactions on Robotics , volume=

    CMax-SLAM: Event-Based Rotational-Motion Bundle Adjustment and SLAM System Using Contrast Maximization , author=. IEEE Transactions on Robotics , volume=. 2024 , doi=

    work page 2024

  60. [60]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Event-Based Photometric Bundle Adjustment , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2025 , doi=

    work page 2025

  61. [61]

    IEEE Robotics and Automation Letters , volume=

    Accurate Angular Velocity Estimation With an Event Camera , author=. IEEE Robotics and Automation Letters , volume=. 2017 , doi=

    work page 2017

  62. [62]

    IEEE Transactions on Robotics , volume=

    CMax-SLAM: Event-based Rotational-Motion Bundle Adjustment and SLAM System using Contrast Maximization , author=. IEEE Transactions on Robotics , volume=. 2024 , publisher=

    work page 2024

  63. [63]

    Proceedings of the British Machine Vision Conference (BMVC) , pages=

    Simultaneous Mosaicing and Tracking with an Event Camera , author=. Proceedings of the British Machine Vision Conference (BMVC) , pages=. 2014 , doi=

    work page 2014

  64. [64]

    IEEE International Conference on Computational Photography (ICCP) , pages=

    Real-Time Panoramic Tracking for Event Cameras , author=. IEEE International Conference on Computational Photography (ICCP) , pages=. 2017 , doi=

    work page 2017

  65. [65]

    IEEE Robotics and Automation Letters , volume=

    Real-Time Rotational Motion Estimation With Contrast Maximization Over Globally Aligned Events , author=. IEEE Robotics and Automation Letters , volume=. 2021 , doi=

    work page 2021

  66. [66]

    Sensors , volume=

    Event Collapse in Contrast Maximization Frameworks , author=. Sensors , volume=. 2022 , doi=

    work page 2022

  67. [67]

    IEEE Transactions on Image Processing , volume=

    Unsupervised Deep Image Stitching: Reconstructing Stitched Features to Images , author=. IEEE Transactions on Image Processing , volume=. 2021 , doi=

    work page 2021

  68. [68]

    Scientific Reports , volume=

    Image Rectangling Network Based on Reparameterized Transformer and Assisted Learning , author=. Scientific Reports , volume=. 2024 , doi=

    work page 2024

  69. [69]

    arXiv preprint arXiv:2501.xxxxx , year=

    PIS3R: Very large parallax image stitching via deep 3D reconstruction , author=. arXiv preprint arXiv:2501.xxxxx , year=

    work page

  70. [70]

    arXiv preprint arXiv:2311.xxxxx , year=

    SandFall: A deep learning-based image stitching method for unstructured camera arrays , author=. arXiv preprint arXiv:2311.xxxxx , year=

    work page

  71. [71]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2018 , doi=

    work page 2018

  72. [72]

    International Conference on Learning Representations (ICLR) , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations (ICLR) , year=

  73. [73]

    International Conference on Learning Representations (ICLR) , year=

    SGDR: Stochastic Gradient Descent with Warm Restarts , author=. International Conference on Learning Representations (ICLR) , year=

    work page

  74. [74]

    BAUTECHNIK , volume=

    Automatic stitching of fragmented construction plans of hydraulic structures , author=. BAUTECHNIK , volume=. 2022 , publisher=

    work page 2022

  75. [75]

    IEEE Transactions on Image Processing , volume=

    No-Reference Image Quality Assessment in the Spatial Domain , author=. IEEE Transactions on Image Processing , volume=. 2012 , doi=

    work page 2012

  76. [76]

    International Conference on Learning Representations (ICLR) , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations (ICLR) , year=

  77. [77]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages=

    CED: Color Event Camera Dataset , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages=. 2019 , doi=

    work page 2019

  78. [78]

    IEEE Transactions on Information Theory , volume=

    Divergence Measures Based on the Shannon Entropy , author=. IEEE Transactions on Information Theory , volume=. 1991 , doi=

    work page 1991

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.